Merge branch 'develop' of github.com:PaddlePaddle/DeepSpeech into merge_parakeet

deepspeech/__init__.py

@@ -233,7 +233,8 @@ def is_broadcastable(shp1, shp2):
 def masked_fill(xs: paddle.Tensor,
                 mask: paddle.Tensor,
                 value: Union[float, int]):
-    assert is_broadcastable(xs.shape, mask.shape) is True
+    assert is_broadcastable(xs.shape, mask.shape) is True, (xs.shape,
+                                                            mask.shape)
     bshape = paddle.broadcast_shape(xs.shape, mask.shape)
     mask = mask.broadcast_to(bshape)
     trues = paddle.ones_like(xs) * value
@@ -315,7 +316,7 @@ def to(x: paddle.Tensor, *args, **kwargs) -> paddle.Tensor:
     assert len(args) == 1
     if isinstance(args[0], str):  # dtype
         return x.astype(args[0])
-    elif isinstance(args[0], paddle.Tensor):  #Tensor
+    elif isinstance(args[0], paddle.Tensor):  # Tensor
         return x.astype(args[0].dtype)
     else:  # Device
         return x
@@ -354,30 +355,153 @@ if not hasattr(paddle.Tensor, 'tolist'):
         "register user tolist to paddle.Tensor, remove this when fixed!")
     setattr(paddle.Tensor, 'tolist', tolist)
 
+########### hack paddle.nn #############
+from paddle.nn import Layer
+from typing import Optional
+from typing import Mapping
+from typing import Iterable
+from typing import Tuple
+from typing import Iterator
+from collections import OrderedDict, abc as container_abcs
 
-# hack loss
-def ctc_loss(logits,
-             labels,
-             input_lengths,
-             label_lengths,
-             blank=0,
-             reduction='mean',
-             norm_by_times=True):
-    #logger.info("my ctc loss with norm by times")
-    ## https://github.com/PaddlePaddle/Paddle/blob/f5ca2db2cc/paddle/fluid/operators/warpctc_op.h#L403
-    loss_out = paddle.fluid.layers.warpctc(logits, labels, blank, norm_by_times,
-                                           input_lengths, label_lengths)
-
-    loss_out = paddle.fluid.layers.squeeze(loss_out, [-1])
-    assert reduction in ['mean', 'sum', 'none']
-    if reduction == 'mean':
-        loss_out = paddle.mean(loss_out / label_lengths)
-    elif reduction == 'sum':
-        loss_out = paddle.sum(loss_out)
-    return loss_out
 
+class LayerDict(paddle.nn.Layer):
+    r"""Holds submodules in a dictionary.
 
-logger.debug(
-    "override ctc_loss of paddle.nn.functional if exists, remove this when fixed!"
-)
-F.ctc_loss = ctc_loss
+    :class:`~paddle.nn.LayerDict` can be indexed like a regular Python dictionary,
+    but modules it contains are properly registered, and will be visible by all
+    :class:`~paddle.nn.Layer` methods.
+
+    :class:`~paddle.nn.LayerDict` is an **ordered** dictionary that respects
+
+    * the order of insertion, and
+
+    * in :meth:`~paddle.nn.LayerDict.update`, the order of the merged
+      ``OrderedDict``, ``dict`` (started from Python 3.6) or another
+      :class:`~paddle.nn.LayerDict` (the argument to
+      :meth:`~paddle.nn.LayerDict.update`).
+
+    Note that :meth:`~paddle.nn.LayerDict.update` with other unordered mapping
+    types (e.g., Python's plain ``dict`` before Python version 3.6) does not
+    preserve the order of the merged mapping.
+
+    Args:
+        modules (iterable, optional): a mapping (dictionary) of (string: module)
+            or an iterable of key-value pairs of type (string, module)
+
+    Example::
+
+        class MyModule(nn.Layer):
+            def __init__(self):
+                super(MyModule, self).__init__()
+                self.choices = nn.LayerDict({
+                        'conv': nn.Conv2d(10, 10, 3),
+                        'pool': nn.MaxPool2d(3)
+                })
+                self.activations = nn.LayerDict([
+                        ['lrelu', nn.LeakyReLU()],
+                        ['prelu', nn.PReLU()]
+                ])
+
+            def forward(self, x, choice, act):
+                x = self.choices[choice](x)
+                x = self.activations[act](x)
+                return x
+    """
+
+    def __init__(self, modules: Optional[Mapping[str, Layer]]=None) -> None:
+        super(LayerDict, self).__init__()
+        if modules is not None:
+            self.update(modules)
+
+    def __getitem__(self, key: str) -> Layer:
+        return self._modules[key]
+
+    def __setitem__(self, key: str, module: Layer) -> None:
+        self.add_module(key, module)
+
+    def __delitem__(self, key: str) -> None:
+        del self._modules[key]
+
+    def __len__(self) -> int:
+        return len(self._modules)
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(self._modules)
+
+    def __contains__(self, key: str) -> bool:
+        return key in self._modules
+
+    def clear(self) -> None:
+        """Remove all items from the LayerDict.
+        """
+        self._modules.clear()
+
+    def pop(self, key: str) -> Layer:
+        r"""Remove key from the LayerDict and return its module.
+
+        Args:
+            key (string): key to pop from the LayerDict
+        """
+        v = self[key]
+        del self[key]
+        return v
+
+    def keys(self) -> Iterable[str]:
+        r"""Return an iterable of the LayerDict keys.
+        """
+        return self._modules.keys()
+
+    def items(self) -> Iterable[Tuple[str, Layer]]:
+        r"""Return an iterable of the LayerDict key/value pairs.
+        """
+        return self._modules.items()
+
+    def values(self) -> Iterable[Layer]:
+        r"""Return an iterable of the LayerDict values.
+        """
+        return self._modules.values()
+
+    def update(self, modules: Mapping[str, Layer]) -> None:
+        r"""Update the :class:`~paddle.nn.LayerDict` with the key-value pairs from a
+        mapping or an iterable, overwriting existing keys.
+
+        .. note::
+            If :attr:`modules` is an ``OrderedDict``, a :class:`~paddle.nn.LayerDict`, or
+            an iterable of key-value pairs, the order of new elements in it is preserved.
+
+        Args:
+            modules (iterable): a mapping (dictionary) from string to :class:`~paddle.nn.Layer`,
+                or an iterable of key-value pairs of type (string, :class:`~paddle.nn.Layer`)
+        """
+        if not isinstance(modules, container_abcs.Iterable):
+            raise TypeError("LayerDict.update should be called with an "
+                            "iterable of key/value pairs, but got " + type(
+                                modules).__name__)
+
+        if isinstance(modules,
+                      (OrderedDict, LayerDict, container_abcs.Mapping)):
+            for key, module in modules.items():
+                self[key] = module
+        else:
+            # modules here can be a list with two items
+            for j, m in enumerate(modules):
+                if not isinstance(m, container_abcs.Iterable):
+                    raise TypeError("LayerDict update sequence element "
+                                    "#" + str(j) + " should be Iterable; is" +
+                                    type(m).__name__)
+                if not len(m) == 2:
+                    raise ValueError("LayerDict update sequence element "
+                                     "#" + str(j) + " has length " + str(
+                                         len(m)) + "; 2 is required")
+                # modules can be Mapping (what it's typed at), or a list: [(name1, module1), (name2, module2)]
+                # that's too cumbersome to type correctly with overloads, so we add an ignore here
+                self[m[0]] = m[1]  # type: ignore[assignment]
+
+    # remove forward alltogether to fallback on Module's _forward_unimplemented
+
+
+if not hasattr(paddle.nn, 'LayerDict'):
+    logger.debug(
+        "register user LayerDict to paddle.nn, remove this when fixed!")
+    setattr(paddle.nn, 'LayerDict', LayerDict)

deepspeech/decoders/beam_search/__init__.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from .batch_beam_search import BatchBeamSearch
+from .beam_search import beam_search
+from .beam_search import BeamSearch
+from .beam_search import Hypothesis

deepspeech/decoders/beam_search/batch_beam_search.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+class BatchBeamSearch():
+    pass

deepspeech/decoders/ctcdecoder/swig/ctc_greedy_decoder.cpp

@@ -53,7 +53,8 @@ std::string ctc_greedy_decoder(
     std::string best_path_result;
     for (size_t i = 0; i < idx_vec.size(); ++i) {
         if (idx_vec[i] != blank_id) {
-            best_path_result += vocabulary[idx_vec[i]];
+            std::string ch = vocabulary[idx_vec[i]];
+            best_path_result += (ch == kSPACE) ? tSPACE : ch;
         }
     }
     return best_path_result;

deepspeech/decoders/ctcdecoder/swig/decoder_utils.cpp

@@ -74,7 +74,8 @@ std::vector<std::pair<double, std::string>> get_beam_search_result(
         // convert index to string
         std::string output_str;
         for (size_t j = 0; j < output.size(); j++) {
-            output_str += vocabulary[output[j]];
+            std::string ch = vocabulary[output[j]];
+            output_str += (ch == kSPACE) ? tSPACE : ch;
         }
         std::pair<double, std::string> output_pair(
             -space_prefixes[i]->approx_ctc, output_str);

deepspeech/decoders/ctcdecoder/swig/decoder_utils.h

@@ -21,6 +21,7 @@
 #include "path_trie.h"
 
 const std::string kSPACE = "<space>";
+const std::string tSPACE = " ";
 const float NUM_FLT_INF = std::numeric_limits<float>::max();
 const float NUM_FLT_MIN = std::numeric_limits<float>::min();
 

deepspeech/decoders/recog.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""V2 backend for `asr_recog.py` using py:class:`decoders.beam_search.BeamSearch`."""
+import jsonlines
+import paddle
+from yacs.config import CfgNode
+
+from .beam_search import BatchBeamSearch
+from .beam_search import BeamSearch
+from .scorers.length_bonus import LengthBonus
+from .scorers.scorer_interface import BatchScorerInterface
+from .utils import add_results_to_json
+from deepspeech.exps import dynamic_import_tester
+from deepspeech.io.reader import LoadInputsAndTargets
+from deepspeech.models.asr_interface import ASRInterface
+from deepspeech.utils.log import Log
+# from espnet.asr.asr_utils import get_model_conf
+# from espnet.asr.asr_utils import torch_load
+# from espnet.nets.lm_interface import dynamic_import_lm
+
+logger = Log(__name__).getlog()
+
+# NOTE: you need this func to generate our sphinx doc
+
+
+def load_trained_model(args):
+    args.nprocs = args.ngpu
+    confs = CfgNode()
+    confs.set_new_allowed(True)
+    confs.merge_from_file(args.model_conf)
+    class_obj = dynamic_import_tester(args.model_name)
+    exp = class_obj(confs, args)
+    with exp.eval():
+        exp.setup()
+        exp.restore()
+    char_list = exp.args.char_list
+    model = exp.model
+    return model, char_list, exp, confs
+
+
+def recog_v2(args):
+    """Decode with custom models that implements ScorerInterface.
+
+    Args:
+        args (namespace): The program arguments.
+        See py:func:`bin.asr_recog.get_parser` for details
+
+    """
+    logger.warning("experimental API for custom LMs is selected by --api v2")
+    if args.batchsize > 1:
+        raise NotImplementedError("multi-utt batch decoding is not implemented")
+    if args.streaming_mode is not None:
+        raise NotImplementedError("streaming mode is not implemented")
+    if args.word_rnnlm:
+        raise NotImplementedError("word LM is not implemented")
+
+    # set_deterministic(args)
+    model, char_list, exp, confs = load_trained_model(args)
+    assert isinstance(model, ASRInterface)
+
+    load_inputs_and_targets = LoadInputsAndTargets(
+        mode="asr",
+        load_output=False,
+        sort_in_input_length=False,
+        preprocess_conf=confs.collator.augmentation_config
+        if args.preprocess_conf is None else args.preprocess_conf,
+        preprocess_args={"train": False}, )
+
+    if args.rnnlm:
+        lm_args = get_model_conf(args.rnnlm, args.rnnlm_conf)
+        # NOTE: for a compatibility with less than 0.5.0 version models
+        lm_model_module = getattr(lm_args, "model_module", "default")
+        lm_class = dynamic_import_lm(lm_model_module, lm_args.backend)
+        lm = lm_class(len(char_list), lm_args)
+        torch_load(args.rnnlm, lm)
+        lm.eval()
+    else:
+        lm = None
+
+    if args.ngram_model:
+        from .scorers.ngram import NgramFullScorer
+        from .scorers.ngram import NgramPartScorer
+
+        if args.ngram_scorer == "full":
+            ngram = NgramFullScorer(args.ngram_model, char_list)
+        else:
+            ngram = NgramPartScorer(args.ngram_model, char_list)
+    else:
+        ngram = None
+
+    scorers = model.scorers()  # decoder
+    scorers["lm"] = lm
+    scorers["ngram"] = ngram
+    scorers["length_bonus"] = LengthBonus(len(char_list))
+    weights = dict(
+        decoder=1.0 - args.ctc_weight,
+        ctc=args.ctc_weight,
+        lm=args.lm_weight,
+        ngram=args.ngram_weight,
+        length_bonus=args.penalty, )
+    beam_search = BeamSearch(
+        beam_size=args.beam_size,
+        vocab_size=len(char_list),
+        weights=weights,
+        scorers=scorers,
+        sos=model.sos,
+        eos=model.eos,
+        token_list=char_list,
+        pre_beam_score_key=None if args.ctc_weight == 1.0 else "full", )
+
+    # TODO(karita): make all scorers batchfied
+    if args.batchsize == 1:
+        non_batch = [
+            k for k, v in beam_search.full_scorers.items()
+            if not isinstance(v, BatchScorerInterface)
+        ]
+        if len(non_batch) == 0:
+            beam_search.__class__ = BatchBeamSearch
+            logger.info("BatchBeamSearch implementation is selected.")
+        else:
+            logger.warning(f"As non-batch scorers {non_batch} are found, "
+                           f"fall back to non-batch implementation.")
+
+    if args.ngpu > 1:
+        raise NotImplementedError("only single GPU decoding is supported")
+    if args.ngpu == 1:
+        device = "gpu:0"
+    else:
+        device = "cpu"
+    paddle.set_device(device)
+    dtype = getattr(paddle, args.dtype)
+    logger.info(f"Decoding device={device}, dtype={dtype}")
+    model.to(device=device, dtype=dtype)
+    model.eval()
+    beam_search.to(device=device, dtype=dtype)
+    beam_search.eval()
+
+    # read json data
+    js = []
+    with jsonlines.open(args.recog_json, "r") as reader:
+        for item in reader:
+            js.append(item)
+    # jsonlines to dict, key by 'utt', value by jsonline
+    js = {item['utt']: item for item in js}
+
+    new_js = {}
+    with paddle.no_grad():
+        with jsonlines.open(args.result_label, "w") as f:
+            for idx, name in enumerate(js.keys(), 1):
+                logger.info(f"({idx}/{len(js.keys())}) decoding " + name)
+                batch = [(name, js[name])]
+                feat = load_inputs_and_targets(batch)[0][0]
+                logger.info(f'feat: {feat.shape}')
+                enc = model.encode(paddle.to_tensor(feat).to(dtype))
+                logger.info(f'eout: {enc.shape}')
+                nbest_hyps = beam_search(
+                    x=enc,
+                    maxlenratio=args.maxlenratio,
+                    minlenratio=args.minlenratio)
+                nbest_hyps = [
+                    h.asdict()
+                    for h in nbest_hyps[:min(len(nbest_hyps), args.nbest)]
+                ]
+                new_js[name] = add_results_to_json(js[name], nbest_hyps,
+                                                   char_list)
+
+                item = new_js[name]['output'][0]  # 1-best
+                ref = item['text']
+                rec_text = item['rec_text'].replace('▁', ' ').replace(
+                    '<eos>', '').strip()
+                rec_tokenid = list(map(int, item['rec_tokenid'].split()))
+                f.write({
+                    "utt": name,
+                    "refs": [ref],
+                    "hyps": [rec_text],
+                    "hyps_tokenid": [rec_tokenid],
+                })

deepspeech/decoders/recog_bin.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""End-to-end speech recognition model decoding script."""
+import logging
+import os
+import random
+import sys
+from distutils.util import strtobool
+
+import configargparse
+import numpy as np
+
+from deepspeech.decoders.recog import recog_v2
+
+
+def get_parser():
+    """Get default arguments."""
+    parser = configargparse.ArgumentParser(
+        description="Transcribe text from speech using "
+        "a speech recognition model on one CPU or GPU",
+        config_file_parser_class=configargparse.YAMLConfigFileParser,
+        formatter_class=configargparse.ArgumentDefaultsHelpFormatter, )
+    parser.add(
+        '--model-name',
+        type=str,
+        default='u2_kaldi',
+        help='model name, e.g: deepspeech2, u2, u2_kaldi, u2_st')
+    # general configuration
+    parser.add("--config", is_config_file=True, help="Config file path")
+    parser.add(
+        "--config2",
+        is_config_file=True,
+        help="Second config file path that overwrites the settings in `--config`",
+    )
+    parser.add(
+        "--config3",
+        is_config_file=True,
+        help="Third config file path that overwrites the settings "
+        "in `--config` and `--config2`", )
+
+    parser.add_argument("--ngpu", type=int, default=0, help="Number of GPUs")
+    parser.add_argument(
+        "--dtype",
+        choices=("float16", "float32", "float64"),
+        default="float32",
+        help="Float precision (only available in --api v2)", )
+    parser.add_argument("--debugmode", type=int, default=1, help="Debugmode")
+    parser.add_argument("--seed", type=int, default=1, help="Random seed")
+    parser.add_argument(
+        "--verbose", "-V", type=int, default=2, help="Verbose option")
+    parser.add_argument(
+        "--batchsize",
+        type=int,
+        default=1,
+        help="Batch size for beam search (0: means no batch processing)", )
+    parser.add_argument(
+        "--preprocess-conf",
+        type=str,
+        default=None,
+        help="The configuration file for the pre-processing", )
+    parser.add_argument(
+        "--api",
+        default="v2",
+        choices=["v2"],
+        help="Beam search APIs "
+        "v2: Experimental API. It supports any models that implements ScorerInterface.",
+    )
+    # task related
+    parser.add_argument(
+        "--recog-json", type=str, help="Filename of recognition data (json)")
+    parser.add_argument(
+        "--result-label",
+        type=str,
+        required=True,
+        help="Filename of result label data (json)", )
+    # model (parameter) related
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Model file parameters to read")
+    parser.add_argument(
+        "--model-conf", type=str, default=None, help="Model config file")
+    parser.add_argument(
+        "--num-spkrs",
+        type=int,
+        default=1,
+        choices=[1, 2],
+        help="Number of speakers in the speech", )
+    parser.add_argument(
+        "--num-encs",
+        default=1,
+        type=int,
+        help="Number of encoders in the model.")
+    # search related
+    parser.add_argument(
+        "--nbest", type=int, default=1, help="Output N-best hypotheses")
+    parser.add_argument("--beam-size", type=int, default=1, help="Beam size")
+    parser.add_argument(
+        "--penalty", type=float, default=0.0, help="Incertion penalty")
+    parser.add_argument(
+        "--maxlenratio",
+        type=float,
+        default=0.0,
+        help="""Input length ratio to obtain max output length.
+                        If maxlenratio=0.0 (default), it uses a end-detect function
+                        to automatically find maximum hypothesis lengths.
+                        If maxlenratio<0.0, its absolute value is interpreted
+                        as a constant max output length""", )
+    parser.add_argument(
+        "--minlenratio",
+        type=float,
+        default=0.0,
+        help="Input length ratio to obtain min output length", )
+    parser.add_argument(
+        "--ctc-weight",
+        type=float,
+        default=0.0,
+        help="CTC weight in joint decoding")
+    parser.add_argument(
+        "--weights-ctc-dec",
+        type=float,
+        action="append",
+        help="ctc weight assigned to each encoder during decoding."
+        "[in multi-encoder mode only]", )
+    parser.add_argument(
+        "--ctc-window-margin",
+        type=int,
+        default=0,
+        help="""Use CTC window with margin parameter to accelerate
+                        CTC/attention decoding especially on GPU. Smaller magin
+                        makes decoding faster, but may increase search errors.
+                        If margin=0 (default), this function is disabled""", )
+    # transducer related
+    parser.add_argument(
+        "--search-type",
+        type=str,
+        default="default",
+        choices=["default", "nsc", "tsd", "alsd", "maes"],
+        help="""Type of beam search implementation to use during inference.
+        Can be either: default beam search ("default"),
+        N-Step Constrained beam search ("nsc"), Time-Synchronous Decoding ("tsd"),
+        Alignment-Length Synchronous Decoding ("alsd") or
+        modified Adaptive Expansion Search ("maes").""", )
+    parser.add_argument(
+        "--nstep",
+        type=int,
+        default=1,
+        help="""Number of expansion steps allowed in NSC beam search or mAES
+        (nstep > 0 for NSC and nstep > 1 for mAES).""", )
+    parser.add_argument(
+        "--prefix-alpha",
+        type=int,
+        default=2,
+        help="Length prefix difference allowed in NSC beam search or mAES.", )
+    parser.add_argument(
+        "--max-sym-exp",
+        type=int,
+        default=2,
+        help="Number of symbol expansions allowed in TSD.", )
+    parser.add_argument(
+        "--u-max",
+        type=int,
+        default=400,
+        help="Length prefix difference allowed in ALSD.", )
+    parser.add_argument(
+        "--expansion-gamma",
+        type=float,
+        default=2.3,
+        help="Allowed logp difference for prune-by-value method in mAES.", )
+    parser.add_argument(
+        "--expansion-beta",
+        type=int,
+        default=2,
+        help="""Number of additional candidates for expanded hypotheses
+                selection in mAES.""", )
+    parser.add_argument(
+        "--score-norm",
+        type=strtobool,
+        nargs="?",
+        default=True,
+        help="Normalize final hypotheses' score by length", )
+    parser.add_argument(
+        "--softmax-temperature",
+        type=float,
+        default=1.0,
+        help="Penalization term for softmax function.", )
+    # rnnlm related
+    parser.add_argument(
+        "--rnnlm", type=str, default=None, help="RNNLM model file to read")
+    parser.add_argument(
+        "--rnnlm-conf",
+        type=str,
+        default=None,
+        help="RNNLM model config file to read")
+    parser.add_argument(
+        "--word-rnnlm",
+        type=str,
+        default=None,
+        help="Word RNNLM model file to read")
+    parser.add_argument(
+        "--word-rnnlm-conf",
+        type=str,
+        default=None,
+        help="Word RNNLM model config file to read", )
+    parser.add_argument(
+        "--word-dict", type=str, default=None, help="Word list to read")
+    parser.add_argument(
+        "--lm-weight", type=float, default=0.1, help="RNNLM weight")
+    # ngram related
+    parser.add_argument(
+        "--ngram-model",
+        type=str,
+        default=None,
+        help="ngram model file to read")
+    parser.add_argument(
+        "--ngram-weight", type=float, default=0.1, help="ngram weight")
+    parser.add_argument(
+        "--ngram-scorer",
+        type=str,
+        default="part",
+        choices=("full", "part"),
+        help="""if the ngram is set as a part scorer, similar with CTC scorer,
+                ngram scorer only scores topK hypethesis.
+                if the ngram is set as full scorer, ngram scorer scores all hypthesis
+                the decoding speed of part scorer is musch faster than full one""",
+    )
+    # streaming related
+    parser.add_argument(
+        "--streaming-mode",
+        type=str,
+        default=None,
+        choices=["window", "segment"],
+        help="""Use streaming recognizer for inference.
+                        `--batchsize` must be set to 0 to enable this mode""", )
+    parser.add_argument(
+        "--streaming-window", type=int, default=10, help="Window size")
+    parser.add_argument(
+        "--streaming-min-blank-dur",
+        type=int,
+        default=10,
+        help="Minimum blank duration threshold", )
+    parser.add_argument(
+        "--streaming-onset-margin", type=int, default=1, help="Onset margin")
+    parser.add_argument(
+        "--streaming-offset-margin", type=int, default=1, help="Offset margin")
+    # non-autoregressive related
+    # Mask CTC related. See https://arxiv.org/abs/2005.08700 for the detail.
+    parser.add_argument(
+        "--maskctc-n-iterations",
+        type=int,
+        default=10,
+        help="Number of decoding iterations."
+        "For Mask CTC, set 0 to predict 1 mask/iter.", )
+    parser.add_argument(
+        "--maskctc-probability-threshold",
+        type=float,
+        default=0.999,
+        help="Threshold probability for CTC output", )
+    # quantize model related
+    parser.add_argument(
+        "--quantize-config",
+        nargs="*",
+        help="Quantize config list. E.g.: --quantize-config=[Linear,LSTM,GRU]",
+    )
+    parser.add_argument(
+        "--quantize-dtype",
+        type=str,
+        default="qint8",
+        help="Dtype dynamic quantize")
+    parser.add_argument(
+        "--quantize-asr-model",
+        type=bool,
+        default=False,
+        help="Quantize asr model", )
+    parser.add_argument(
+        "--quantize-lm-model",
+        type=bool,
+        default=False,
+        help="Quantize lm model", )
+    return parser
+
+
+def main(args):
+    """Run the main decoding function."""
+    parser = get_parser()
+    parser.add_argument(
+        "--output", metavar="CKPT_DIR", help="path to save checkpoint.")
+    parser.add_argument(
+        "--checkpoint_path", type=str, help="path to load checkpoint")
+    parser.add_argument("--dict-path", type=str, help="path to load checkpoint")
+    args = parser.parse_args(args)
+
+    if args.ngpu == 0 and args.dtype == "float16":
+        raise ValueError(
+            f"--dtype {args.dtype} does not support the CPU backend.")
+
+    # logging info
+    if args.verbose == 1:
+        logging.basicConfig(
+            level=logging.INFO,
+            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
+        )
+    elif args.verbose == 2:
+        logging.basicConfig(
+            level=logging.DEBUG,
+            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
+        )
+    else:
+        logging.basicConfig(
+            level=logging.WARN,
+            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
+        )
+        logging.warning("Skip DEBUG/INFO messages")
+    logging.info(args)
+
+    # check CUDA_VISIBLE_DEVICES
+    if args.ngpu > 0:
+        cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
+        if cvd is None:
+            logging.warning("CUDA_VISIBLE_DEVICES is not set.")
+        elif args.ngpu != len(cvd.split(",")):
+            logging.error("#gpus is not matched with CUDA_VISIBLE_DEVICES.")
+            sys.exit(1)
+
+        # TODO(mn5k): support of multiple GPUs
+        if args.ngpu > 1:
+            logging.error("The program only supports ngpu=1.")
+            sys.exit(1)
+
+    # display PYTHONPATH
+    logging.info("python path = " + os.environ.get("PYTHONPATH", "(None)"))
+
+    # seed setting
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    logging.info("set random seed = %d" % args.seed)
+
+    # validate rnn options
+    if args.rnnlm is not None and args.word_rnnlm is not None:
+        logging.error(
+            "It seems that both --rnnlm and --word-rnnlm are specified. "
+            "Please use either option.")
+        sys.exit(1)
+
+    # recog
+    if args.num_spkrs == 1:
+        if args.num_encs == 1:
+            # Experimental API that supports custom LMs
+            if args.api == "v2":
+
+                recog_v2(args)
+            else:
+                raise ValueError("Only support --api v2")
+        else:
+            if args.api == "v2":
+                raise NotImplementedError(
+                    f"--num-encs {args.num_encs} > 1 is not supported in --api v2"
+                )
+    elif args.num_spkrs == 2:
+        raise ValueError("asr_mix not supported.")
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])

deepspeech/decoders/scores/score_interface.py → deepspeech/decoders/scorers/scorer_interface.py

@@ -145,9 +145,11 @@ class PartialScorerInterface(ScorerInterface):
     and receives pre-pruned next tokens to score because it is too heavy to score
     all the tokens.
 
+    Score sub-set of tokens, not all.
+
     Examples:
          * Prefix search for connectionist-temporal-classification models
-             * :class:`espnet.nets.scorers.ctc.CTCPrefixScorer`
+             * :class:`decoders.scorers.ctc.CTCPrefixScorer`
 
     """
 

deepspeech/decoders/utils.py

@@ -11,8 +11,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import numpy as np
 
-__all__ = ["end_detect"]
+from deepspeech.utils.log import Log
+logger = Log(__name__).getlog()
+
+__all__ = ["end_detect", "parse_hypothesis", "add_results_to_json"]
 
 
 def end_detect(ended_hyps, i, M=3, D_end=np.log(1 * np.exp(-10))):
@@ -47,3 +51,79 @@ def end_detect(ended_hyps, i, M=3, D_end=np.log(1 * np.exp(-10))):
         return True
     else:
         return False
+
+
+# * ------------------ recognition related ------------------ *
+def parse_hypothesis(hyp, char_list):
+    """Parse hypothesis.
+
+    Args:
+        hyp (list[dict[str, Any]]): Recognition hypothesis.
+        char_list (list[str]): List of characters.
+
+    Returns:
+        tuple(str, str, str, float)
+
+    """
+    # remove sos and get results
+    tokenid_as_list = list(map(int, hyp["yseq"][1:]))
+    token_as_list = [char_list[idx] for idx in tokenid_as_list]
+    score = float(hyp["score"])
+
+    # convert to string
+    tokenid = " ".join([str(idx) for idx in tokenid_as_list])
+    token = " ".join(token_as_list)
+    text = "".join(token_as_list).replace("<space>", " ")
+
+    return text, token, tokenid, score
+
+
+def add_results_to_json(js, nbest_hyps, char_list):
+    """Add N-best results to json.
+
+    Args:
+        js (dict[str, Any]): Groundtruth utterance dict.
+        nbest_hyps_sd (list[dict[str, Any]]):
+            List of hypothesis for multi_speakers: nutts x nspkrs.
+        char_list (list[str]): List of characters.
+
+    Returns:
+        dict[str, Any]: N-best results added utterance dict.
+
+    """
+    # copy old json info
+    new_js = dict()
+    new_js["utt2spk"] = js["utt2spk"]
+    new_js["output"] = []
+
+    for n, hyp in enumerate(nbest_hyps, 1):
+        # parse hypothesis
+        rec_text, rec_token, rec_tokenid, score = parse_hypothesis(hyp,
+                                                                   char_list)
+
+        # copy ground-truth
+        if len(js["output"]) > 0:
+            out_dic = dict(js["output"][0].items())
+        else:
+            # for no reference case (e.g., speech translation)
+            out_dic = {"name": ""}
+
+        # update name
+        out_dic["name"] += "[%d]" % n
+
+        # add recognition results
+        out_dic["rec_text"] = rec_text
+        out_dic["rec_token"] = rec_token
+        out_dic["rec_tokenid"] = rec_tokenid
+        out_dic["score"] = score
+
+        # add to list of N-best result dicts
+        new_js["output"].append(out_dic)
+
+        # show 1-best result
+        if n == 1:
+            if "text" in out_dic.keys():
+                logger.info("groundtruth: %s" % out_dic["text"])
+            logger.info("prediction : %s" % out_dic["rec_text"])
+
+    return new_js

deepspeech/exps/__init__.py

@@ -11,3 +11,52 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from deepspeech.training.trainer import Trainer
+from deepspeech.utils.dynamic_import import dynamic_import
+
+model_trainer_alias = {
+    "ds2": "deepspeech.exp.deepspeech2.model:DeepSpeech2Trainer",
+    "u2": "deepspeech.exps.u2.model:U2Trainer",
+    "u2_kaldi": "deepspeech.exps.u2_kaldi.model:U2Trainer",
+    "u2_st": "deepspeech.exps.u2_st.model:U2STTrainer",
+}
+
+
+def dynamic_import_trainer(module):
+    """Import Trainer dynamically.
+
+    Args:
+        module (str): trainer name. e.g., ds2, u2, u2_kaldi
+
+    Returns:
+        type: Trainer class
+
+    """
+    model_class = dynamic_import(module, model_trainer_alias)
+    assert issubclass(model_class,
+                      Trainer), f"{module} does not implement Trainer"
+    return model_class
+
+
+model_tester_alias = {
+    "ds2": "deepspeech.exp.deepspeech2.model:DeepSpeech2Tester",
+    "u2": "deepspeech.exps.u2.model:U2Tester",
+    "u2_kaldi": "deepspeech.exps.u2_kaldi.model:U2Tester",
+    "u2_st": "deepspeech.exps.u2_st.model:U2STTester",
+}
+
+
+def dynamic_import_tester(module):
+    """Import Tester dynamically.
+
+    Args:
+        module (str): tester name. e.g., ds2, u2, u2_kaldi
+
+    Returns:
+        type: Tester class
+
+    """
+    model_class = dynamic_import(module, model_tester_alias)
+    assert issubclass(model_class,
+                      Trainer), f"{module} does not implement Tester"
+    return model_class

deepspeech/exps/deepspeech2/bin/test_hub.py

@@ -56,11 +56,6 @@ class DeepSpeech2Tester_hub():
             cutoff_prob=cfg.cutoff_prob,
             cutoff_top_n=cfg.cutoff_top_n,
             num_processes=cfg.num_proc_bsearch)
-        #replace the '<space>' with ' '
-        result_transcripts = [
-            self._text_featurizer.detokenize(sentence)
-            for sentence in result_transcripts
-        ]
 
         return result_transcripts
 
@@ -167,9 +162,8 @@ def check(audio_file):
 
 def main_sp(config, args):
     exp = DeepSpeech2Tester_hub(config, args)
-    with exp.eval():
-        exp.setup()
-        exp.run_test()
+    exp.setup()
+    exp.run_test()
 
 
 def main(config, args):

deepspeech/exps/deepspeech2/model.py

@@ -341,11 +341,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             cutoff_prob=cfg.cutoff_prob,
             cutoff_top_n=cfg.cutoff_top_n,
             num_processes=cfg.num_proc_bsearch)
-        #replace the <space> with ' '
-        result_transcripts = [
-            self._text_featurizer.detokenize(sentence)
-            for sentence in result_transcripts
-        ]
 
         self.autolog.times.stamp()
         self.autolog.times.stamp()

deepspeech/exps/u2/bin/test_hub.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Evaluation for U2 model."""
+import cProfile
+import os
+import sys
+
+import paddle
+import soundfile
+
+from deepspeech.exps.u2.config import get_cfg_defaults
+from deepspeech.frontend.featurizer.text_featurizer import TextFeaturizer
+from deepspeech.io.collator import SpeechCollator
+from deepspeech.models.u2 import U2Model
+from deepspeech.training.cli import default_argument_parser
+from deepspeech.training.trainer import Trainer
+from deepspeech.utils import layer_tools
+from deepspeech.utils import mp_tools
+from deepspeech.utils.log import Log
+from deepspeech.utils.utility import print_arguments
+from deepspeech.utils.utility import UpdateConfig
+logger = Log(__name__).getlog()
+
+# TODO(hui zhang): dynamic load
+
+
+class U2Tester_Hub(Trainer):
+    def __init__(self, config, args):
+        # super().__init__(config, args)
+        self.args = args
+        self.config = config
+        self.audio_file = args.audio_file
+        self.collate_fn_test = SpeechCollator.from_config(config)
+        self._text_featurizer = TextFeaturizer(
+            unit_type=config.collator.unit_type,
+            vocab_filepath=None,
+            spm_model_prefix=config.collator.spm_model_prefix)
+
+    def setup_model(self):
+        config = self.config
+        model_conf = config.model
+
+        with UpdateConfig(model_conf):
+            model_conf.input_dim = self.collate_fn_test.feature_size
+            model_conf.output_dim = self.collate_fn_test.vocab_size
+
+        model = U2Model.from_config(model_conf)
+
+        if self.parallel:
+            model = paddle.DataParallel(model)
+
+        logger.info(f"{model}")
+        layer_tools.print_params(model, logger.info)
+
+        self.model = model
+        logger.info("Setup model")
+
+    @mp_tools.rank_zero_only
+    @paddle.no_grad()
+    def test(self):
+        self.model.eval()
+        cfg = self.config.decoding
+        audio_file = self.audio_file
+        collate_fn_test = self.collate_fn_test
+        audio, _ = collate_fn_test.process_utterance(
+            audio_file=audio_file, transcript="Hello")
+        audio_len = audio.shape[0]
+        audio = paddle.to_tensor(audio, dtype='float32')
+        audio_len = paddle.to_tensor(audio_len)
+        audio = paddle.unsqueeze(audio, axis=0)
+        vocab_list = collate_fn_test.vocab_list
+
+        text_feature = self.collate_fn_test.text_feature
+        result_transcripts = self.model.decode(
+            audio,
+            audio_len,
+            text_feature=text_feature,
+            decoding_method=cfg.decoding_method,
+            lang_model_path=cfg.lang_model_path,
+            beam_alpha=cfg.alpha,
+            beam_beta=cfg.beta,
+            beam_size=cfg.beam_size,
+            cutoff_prob=cfg.cutoff_prob,
+            cutoff_top_n=cfg.cutoff_top_n,
+            num_processes=cfg.num_proc_bsearch,
+            ctc_weight=cfg.ctc_weight,
+            decoding_chunk_size=cfg.decoding_chunk_size,
+            num_decoding_left_chunks=cfg.num_decoding_left_chunks,
+            simulate_streaming=cfg.simulate_streaming)
+        logger.info("The result_transcripts: " + result_transcripts[0][0])
+
+    def run_test(self):
+        self.resume()
+        try:
+            self.test()
+        except KeyboardInterrupt:
+            sys.exit(-1)
+
+    def setup(self):
+        """Setup the experiment.
+        """
+        paddle.set_device('gpu' if self.args.nprocs > 0 else 'cpu')
+
+        #self.setup_output_dir()
+        #self.setup_checkpointer()
+
+        #self.setup_dataloader()
+        self.setup_model()
+
+        self.iteration = 0
+        self.epoch = 0
+
+    def resume(self):
+        """Resume from the checkpoint at checkpoints in the output
+        directory or load a specified checkpoint.
+        """
+        params_path = self.args.checkpoint_path + ".pdparams"
+        model_dict = paddle.load(params_path)
+        self.model.set_state_dict(model_dict)
+
+
+def check(audio_file):
+    logger.info("checking the audio file format......")
+    try:
+        sig, sample_rate = soundfile.read(audio_file)
+    except Exception as e:
+        logger.error(str(e))
+        logger.error(
+            "can not open the wav file, please check the audio file format")
+        sys.exit(-1)
+    logger.info("The sample rate is %d" % sample_rate)
+    assert (sample_rate == 16000)
+    logger.info("The audio file format is right")
+
+
+def main_sp(config, args):
+    exp = U2Tester_Hub(config, args)
+    with exp.eval():
+        exp.setup()
+        exp.run_test()
+
+
+def main(config, args):
+    main_sp(config, args)
+
+
+if __name__ == "__main__":
+    parser = default_argument_parser()
+    # save asr result to
+    parser.add_argument(
+        "--result_file", type=str, help="path of save the asr result")
+    parser.add_argument(
+        "--audio_file", type=str, help="path of the input audio file")
+    args = parser.parse_args()
+    print_arguments(args, globals())
+
+    if not os.path.isfile(args.audio_file):
+        print("Please input the right audio file path")
+        sys.exit(-1)
+    check(args.audio_file)
+    # https://yaml.org/type/float.html
+    config = get_cfg_defaults()
+    if args.config:
+        config.merge_from_file(args.config)
+    if args.opts:
+        config.merge_from_list(args.opts)
+    config.freeze()
+    print(config)
+    if args.dump_config:
+        with open(args.dump_config, 'w') as f:
+            print(config, file=f)
+
+    # Setting for profiling
+    pr = cProfile.Profile()
+    pr.runcall(main, config, args)
+    pr.dump_stats('test.profile')

deepspeech/exps/u2_kaldi/bin/recog.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import sys
+
+from deepspeech.decoders.recog_bin import main
+
+if __name__ == "__main__":
+    main(sys.argv[1:])

deepspeech/exps/u2_kaldi/model.py

@@ -317,11 +317,9 @@ class U2Trainer(Trainer):
         with UpdateConfig(model_conf):
             model_conf.input_dim = self.train_loader.feat_dim
             model_conf.output_dim = self.train_loader.vocab_size
-
         model = U2Model.from_config(model_conf)
         if self.parallel:
             model = paddle.DataParallel(model)
-        logger.info(f"{model}")
         layer_tools.print_params(model, logger.info)
 
         # lr

deepspeech/frontend/featurizer/text_featurizer.py

@@ -207,7 +207,7 @@ class TextFeaturizer():
         """Load vocabulary from file."""
         vocab_list = load_dict(vocab_filepath, maskctc)
         assert vocab_list is not None
-        logger.info(f"Vocab: {pformat(vocab_list)}")
+        logger.debug(f"Vocab: {pformat(vocab_list)}")
 
         id2token = dict(
             [(idx, token) for (idx, token) in enumerate(vocab_list)])

deepspeech/io/batchfy.py

@@ -354,7 +354,7 @@ def make_batchset(
     :param int batch_frames_out: maximum number of output frames in a minibatch.
     :param int batch_frames_out: maximum number of input+output frames in a minibatch.
     :param str count: strategy to count maximum size of batch.
-        For choices, see espnet.asr.batchfy.BATCH_COUNT_CHOICES
+        For choices, see io.batchfy.BATCH_COUNT_CHOICES
 
     :param int max_length_in: maximum length of input to decide adaptive batch size
     :param int max_length_out: maximum length of output to decide adaptive batch size

deepspeech/models/asr_interface.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""ASR Interface module."""
+import argparse
+
+from deepspeech.utils.dynamic_import import dynamic_import
+
+
+class ASRInterface:
+    """ASR Interface for ESPnet model implementation."""
+
+    @staticmethod
+    def add_arguments(parser):
+        """Add arguments to parser."""
+        return parser
+
+    @classmethod
+    def build(cls, idim: int, odim: int, **kwargs):
+        """Initialize this class with python-level args.
+
+        Args:
+            idim (int): The number of an input feature dim.
+            odim (int): The number of output vocab.
+
+        Returns:
+            ASRinterface: A new instance of ASRInterface.
+
+        """
+        args = argparse.Namespace(**kwargs)
+        return cls(idim, odim, args)
+
+    def forward(self, xs, ilens, ys, olens):
+        """Compute loss for training.
+
+        :param xs: batch of padded source sequences paddle.Tensor (B, Tmax, idim)
+        :param ilens: batch of lengths of source sequences (B), paddle.Tensor
+        :param ys: batch of padded target sequences paddle.Tensor (B, Lmax)
+        :param olens: batch of lengths of target sequences (B), paddle.Tensor
+        :return: loss value
+        :rtype: paddle.Tensor
+        """
+        raise NotImplementedError("forward method is not implemented")
+
+    def recognize(self, x, recog_args, char_list=None, rnnlm=None):
+        """Recognize x for evaluation.
+
+        :param ndarray x: input acouctic feature (B, T, D) or (T, D)
+        :param namespace recog_args: argment namespace contraining options
+        :param list char_list: list of characters
+        :param paddle.nn.Layer rnnlm: language model module
+        :return: N-best decoding results
+        :rtype: list
+        """
+        raise NotImplementedError("recognize method is not implemented")
+
+    def recognize_batch(self, x, recog_args, char_list=None, rnnlm=None):
+        """Beam search implementation for batch.
+
+        :param paddle.Tensor x: encoder hidden state sequences (B, Tmax, Henc)
+        :param namespace recog_args: argument namespace containing options
+        :param list char_list: list of characters
+        :param paddle.nn.Module rnnlm: language model module
+        :return: N-best decoding results
+        :rtype: list
+        """
+        raise NotImplementedError("Batch decoding is not supported yet.")
+
+    def calculate_all_attentions(self, xs, ilens, ys):
+        """Calculate attention.
+
+        :param list xs: list of padded input sequences [(T1, idim), (T2, idim), ...]
+        :param ndarray ilens: batch of lengths of input sequences (B)
+        :param list ys: list of character id sequence tensor [(L1), (L2), (L3), ...]
+        :return: attention weights (B, Lmax, Tmax)
+        :rtype: float ndarray
+        """
+        raise NotImplementedError(
+            "calculate_all_attentions method is not implemented")
+
+    def calculate_all_ctc_probs(self, xs, ilens, ys):
+        """Calculate CTC probability.
+
+        :param list xs_pad: list of padded input sequences [(T1, idim), (T2, idim), ...]
+        :param ndarray ilens: batch of lengths of input sequences (B)
+        :param list ys: list of character id sequence tensor [(L1), (L2), (L3), ...]
+        :return: CTC probabilities (B, Tmax, vocab)
+        :rtype: float ndarray
+        """
+        raise NotImplementedError(
+            "calculate_all_ctc_probs method is not implemented")
+
+    @property
+    def attention_plot_class(self):
+        """Get attention plot class."""
+        from espnet.asr.asr_utils import PlotAttentionReport
+
+        return PlotAttentionReport
+
+    @property
+    def ctc_plot_class(self):
+        """Get CTC plot class."""
+        from espnet.asr.asr_utils import PlotCTCReport
+
+        return PlotCTCReport
+
+    def get_total_subsampling_factor(self):
+        """Get total subsampling factor."""
+        raise NotImplementedError(
+            "get_total_subsampling_factor method is not implemented")
+
+    def encode(self, feat):
+        """Encode feature in `beam_search` (optional).
+
+        Args:
+            x (numpy.ndarray): input feature (T, D)
+        Returns:
+            paddle.Tensor: encoded feature (T, D)
+        """
+        raise NotImplementedError("encode method is not implemented")
+
+    def scorers(self):
+        """Get scorers for `beam_search` (optional).
+
+        Returns:
+            dict[str, ScorerInterface]: dict of `ScorerInterface` objects
+
+        """
+        raise NotImplementedError("decoders method is not implemented")
+
+
+predefined_asr = {
+    "transformer": "deepspeech.models.u2:U2Model",
+    "conformer": "deepspeech.models.u2:U2Model",
+}
+
+
+def dynamic_import_asr(module):
+    """Import ASR models dynamically.
+
+    Args:
+        module (str): asr name. e.g., transformer, conformer
+
+    Returns:
+        type: ASR class
+
+    """
+    model_class = dynamic_import(module, predefined_asr)
+    assert issubclass(model_class,
+                      ASRInterface), f"{module} does not implement ASRInterface"
+    return model_class

deepspeech/models/u2/u2.py

@@ -28,8 +28,10 @@ from paddle import jit
 from paddle import nn
 from yacs.config import CfgNode
 
+from deepspeech.decoders.scorers.ctc import CTCPrefixScorer
 from deepspeech.frontend.utility import IGNORE_ID
 from deepspeech.frontend.utility import load_cmvn
+from deepspeech.models.asr_interface import ASRInterface
 from deepspeech.modules.cmvn import GlobalCMVN
 from deepspeech.modules.ctc import CTCDecoder
 from deepspeech.modules.decoder import TransformerDecoder
@@ -55,7 +57,7 @@ __all__ = ["U2Model", "U2InferModel"]
 logger = Log(__name__).getlog()
 
 
-class U2BaseModel(nn.Layer):
+class U2BaseModel(ASRInterface, nn.Layer):
     """CTC-Attention hybrid Encoder-Decoder model"""
 
     @classmethod
@@ -120,7 +122,7 @@ class U2BaseModel(nn.Layer):
                  **kwargs):
         assert 0.0 <= ctc_weight <= 1.0, ctc_weight
 
-        super().__init__()
+        nn.Layer.__init__(self)
         # note that eos is the same as sos (equivalent ID)
         self.sos = vocab_size - 1
         self.eos = vocab_size - 1
@@ -813,7 +815,27 @@ class U2BaseModel(nn.Layer):
         return res, res_tokenids
 
 
-class U2Model(U2BaseModel):
+class U2DecodeModel(U2BaseModel):
+    def scorers(self):
+        """Scorers."""
+        return dict(
+            decoder=self.decoder, ctc=CTCPrefixScorer(self.ctc, self.eos))
+
+    def encode(self, x):
+        """Encode acoustic features.
+
+        :param ndarray x: source acoustic feature (T, D)
+        :return: encoder outputs
+        :rtype: paddle.Tensor
+        """
+        self.eval()
+        x = paddle.to_tensor(x).unsqueeze(0)
+        ilen = x.size(1)
+        enc_output, _ = self._forward_encoder(x, ilen)
+        return enc_output.squeeze(0)
+
+
+class U2Model(U2DecodeModel):
     def __init__(self, configs: dict):
         vocab_size, encoder, decoder, ctc = U2Model._init_from_config(configs)
 

deepspeech/modules/ctc.py

@@ -11,6 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from typing import Union
+
 import paddle
 from paddle import nn
 from paddle.nn import functional as F
@@ -40,7 +42,7 @@ class CTCDecoderBase(nn.Layer):
                  dropout_rate: float=0.0,
                  reduction: bool=True,
                  batch_average: bool=True,
-                 grad_norm_type: str="instance"):
+                 grad_norm_type: Union[str, None]=None):
         """CTC decoder
 
         Args:
@@ -49,7 +51,7 @@ class CTCDecoderBase(nn.Layer):
             dropout_rate (float): dropout rate (0.0 ~ 1.0)
             reduction (bool): reduce the CTC loss into a scalar, True for 'sum' or 'none'
             batch_average (bool): do batch dim wise average.
-            grad_norm_type (str): one of 'instance', 'batch', 'frame', None.
+            grad_norm_type (str): Default, None. one of 'instance', 'batch', 'frame', None. 
         """
         assert check_argument_types()
         super().__init__()

deepspeech/modules/decoder.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Decoder definition."""
+from typing import Any
 from typing import List
 from typing import Optional
 from typing import Tuple
@@ -20,10 +21,12 @@ import paddle
 from paddle import nn
 from typeguard import check_argument_types
 
+from deepspeech.decoders.scorers.scorer_interface import BatchScorerInterface
 from deepspeech.modules.attention import MultiHeadedAttention
 from deepspeech.modules.decoder_layer import DecoderLayer
 from deepspeech.modules.embedding import PositionalEncoding
 from deepspeech.modules.mask import make_non_pad_mask
+from deepspeech.modules.mask import make_xs_mask
 from deepspeech.modules.mask import subsequent_mask
 from deepspeech.modules.positionwise_feed_forward import PositionwiseFeedForward
 from deepspeech.utils.log import Log
@@ -33,7 +36,7 @@ logger = Log(__name__).getlog()
 __all__ = ["TransformerDecoder"]
 
 
-class TransformerDecoder(nn.Layer):
+class TransformerDecoder(BatchScorerInterface, nn.Layer):
     """Base class of Transfomer decoder module.
     Args:
         vocab_size: output dim
@@ -71,7 +74,8 @@ class TransformerDecoder(nn.Layer):
             concat_after: bool=False, ):
 
         assert check_argument_types()
-        super().__init__()
+        nn.Layer.__init__(self)
+        self.selfattention_layer_type = 'selfattn'
         attention_dim = encoder_output_size
 
         if input_layer == "embed":
@@ -180,3 +184,63 @@ class TransformerDecoder(nn.Layer):
         if self.use_output_layer:
             y = paddle.log_softmax(self.output_layer(y), axis=-1)
         return y, new_cache
+
+    # beam search API (see ScorerInterface)
+    def score(self, ys, state, x):
+        """Score.
+        ys: (ylen,)
+        x: (xlen, n_feat)
+        """
+        ys_mask = subsequent_mask(len(ys)).unsqueeze(0)  # (B,L,L)
+        x_mask = make_xs_mask(x.unsqueeze(0)).unsqueeze(1)  # (B,1,T)
+        if self.selfattention_layer_type != "selfattn":
+            # TODO(karita): implement cache
+            logging.warning(
+                f"{self.selfattention_layer_type} does not support cached decoding."
+            )
+            state = None
+        logp, state = self.forward_one_step(
+            x.unsqueeze(0), x_mask, ys.unsqueeze(0), ys_mask, cache=state)
+        return logp.squeeze(0), state
+
+    # batch beam search API (see BatchScorerInterface)
+    def batch_score(self,
+                    ys: paddle.Tensor,
+                    states: List[Any],
+                    xs: paddle.Tensor) -> Tuple[paddle.Tensor, List[Any]]:
+        """Score new token batch (required).
+
+        Args:
+            ys (paddle.Tensor): paddle.int64 prefix tokens (n_batch, ylen).
+            states (List[Any]): Scorer states for prefix tokens.
+            xs (paddle.Tensor):
+                The encoder feature that generates ys (n_batch, xlen, n_feat).
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        # merge states
+        n_batch = len(ys)
+        n_layers = len(self.decoders)
+        if states[0] is None:
+            batch_state = None
+        else:
+            # transpose state of [batch, layer] into [layer, batch]
+            batch_state = [
+                paddle.stack([states[b][i] for b in range(n_batch)])
+                for i in range(n_layers)
+            ]
+
+        # batch decoding
+        ys_mask = subsequent_mask(ys.size(-1)).unsqueeze(0)  # (B,L,L)
+        xs_mask = make_xs_mask(xs).unsqueeze(1)  # (B,1,T)
+        logp, states = self.forward_one_step(
+            xs, xs_mask, ys, ys_mask, cache=batch_state)
+
+        # transpose state of [layer, batch] into [batch, layer]
+        state_list = [[states[i][b] for i in range(n_layers)]
+                      for b in range(n_batch)]
+        return logp, state_list

deepspeech/modules/loss.py

@@ -54,7 +54,7 @@ class CTCLoss(nn.Layer):
             self.norm_by_total_logits_len = True
         else:
             raise ValueError(f"CTCLoss Grad Norm no support {grad_norm_type}")
-        self.kwargs = {
+        kwargs = {
             "norm_by_times": self.norm_by_times,
             "norm_by_batchsize": self.norm_by_batchsize,
             "norm_by_total_logits_len": self.norm_by_total_logits_len,
@@ -66,10 +66,9 @@ class CTCLoss(nn.Layer):
         except ValueError:
             # Some function, e.g. built-in function, are failed
             param = {}
-        self._kwargs = {k: v for k, v in self.kwargs.items() if k in param}
-        _notin = {k: v for k, v in self.kwargs.items() if k not in param}
+        self._kwargs = {k: v for k, v in kwargs.items() if k in param}
+        _notin = {k: v for k, v in kwargs.items() if k not in param}
         logger.info(f"{self.loss} kwargs:{self._kwargs}, not support: {_notin}")
-        #self.loss_fn = partial(self.loss.forward, **_kwargs)
 
     def forward(self, logits, ys_pad, hlens, ys_lens):
         """Compute CTC loss.
@@ -89,8 +88,7 @@ class CTCLoss(nn.Layer):
         # logits: (B, L, D) -> (L, B, D)
         logits = logits.transpose([1, 0, 2])
         ys_pad = ys_pad.astype(paddle.int32)
-        #loss = self.loss_fn(logits, ys_pad, hlens, ys_lens)
-        loss = self.loss(logits, ys_pad, hlens, ys_lens)
+        loss = self.loss(logits, ys_pad, hlens, ys_lens, **self._kwargs)
         if self.batch_average:
             # Batch-size average
             loss = loss / B

deepspeech/modules/mask.py

@@ -18,12 +18,25 @@ from deepspeech.utils.log import Log
 logger = Log(__name__).getlog()
 
 __all__ = [
-    "make_pad_mask", "make_non_pad_mask", "subsequent_mask",
+    "make_xs_mask", "make_pad_mask", "make_non_pad_mask", "subsequent_mask",
     "subsequent_chunk_mask", "add_optional_chunk_mask", "mask_finished_scores",
     "mask_finished_preds"
 ]
 
 
+def make_xs_mask(xs: paddle.Tensor, pad_value=0.0) -> paddle.Tensor:
+    """Maks mask tensor containing indices of non-padded part.
+    Args:
+        xs (paddle.Tensor): (B, T, D), zeros for pad.
+    Returns:
+        paddle.Tensor: Mask Tensor indices of non-padded part. (B, T)
+    """
+    pad_frame = paddle.full([1, 1, xs.shape[-1]], pad_value, dtype=xs.dtype)
+    mask = xs != pad_frame
+    mask = mask.all(axis=-1)
+    return mask
+
+
 def make_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
     """Make mask tensor containing indices of padded part.
     See description of make_non_pad_mask.
@@ -31,6 +44,7 @@ def make_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
         lengths (paddle.Tensor): Batch of lengths (B,).
     Returns:
         paddle.Tensor: Mask tensor containing indices of padded part.
+        (B, T)
     Examples:
         >>> lengths = [5, 3, 2]
         >>> make_pad_mask(lengths)
@@ -62,6 +76,7 @@ def make_non_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
         lengths (paddle.Tensor): Batch of lengths (B,).
     Returns:
         paddle.Tensor: mask tensor containing indices of padded part.
+        (B, T)
     Examples:
         >>> lengths = [5, 3, 2]
         >>> make_non_pad_mask(lengths)

deepspeech/training/cli.py

@@ -35,7 +35,7 @@ class LoadFromFile(argparse.Action):
             parser.parse_args(f.read().split(), namespace)
 
 
-def default_argument_parser():
+def default_argument_parser(parser=None):
     r"""A simple yet genral argument parser for experiments with parakeet.
 
     This is used in examples with parakeet. And it is intended to be used by
@@ -62,7 +62,9 @@ def default_argument_parser():
     argparse.ArgumentParser
         the parser
     """
-    parser = argparse.ArgumentParser()
+    if parser is None:
+        parser = argparse.ArgumentParser()
+
     parser.register('action', 'extend', ExtendAction)
     parser.add_argument(
         '--conf', type=open, action=LoadFromFile, help="config file.")

deepspeech/training/trainer.py

@@ -126,7 +126,8 @@ class Trainer():
             logger.info(f"Set seed {args.seed}")
 
         # profiler and benchmark options
-        if self.args.benchmark_batch_size:
+        if hasattr(self.args,
+                   "benchmark_batch_size") and self.args.benchmark_batch_size:
             with UpdateConfig(self.config):
                 self.config.collator.batch_size = self.args.benchmark_batch_size
                 self.config.training.log_interval = 1
@@ -326,12 +327,24 @@ class Trainer():
         finally:
             self.destory()
 
+    def restore(self):
+        """Resume from latest checkpoint at checkpoints in the output
+        directory or load a specified checkpoint.
+
+        If ``args.checkpoint_path`` is not None, load the checkpoint, else
+        resume training.
+        """
+        assert self.args.checkpoint_path
+        infos = self.checkpoint.load_latest_parameters(
+            self.model, checkpoint_path=self.args.checkpoint_path)
+        return infos
+
     def run_test(self):
         """Do Test/Decode"""
         try:
             with Timer("Test/Decode Done: {}"):
                 with self.eval():
-                    self.resume_or_scratch()
+                    self.restore()
                     self.test()
         except KeyboardInterrupt:
             exit(-1)
@@ -341,6 +354,7 @@ class Trainer():
         try:
             with Timer("Export Done: {}"):
                 with self.eval():
+                    self.restore()
                     self.export()
         except KeyboardInterrupt:
             exit(-1)
@@ -350,7 +364,7 @@ class Trainer():
         try:
             with Timer("Align Done: {}"):
                 with self.eval():
-                    self.resume_or_scratch()
+                    self.restore()
                     self.align()
         except KeyboardInterrupt:
             sys.exit(-1)

deepspeech/utils/utility.py

@@ -42,7 +42,7 @@ def all_version():
         "paddle_commit": paddle.version.commit,
         "soundfile": soundfile.__version__,
     }
-    logger.info(f"Deps Module Version:{pformat(vers.items())}")
+    logger.info(f"Deps Module Version:{pformat(list(vers.items()))}")
 
 
 @contextmanager

docs/source/asr/deepspeech_architecture.md

@@ -14,10 +14,11 @@ In addition, the training process and the testing process are also introduced.
 The arcitecture of the model is shown in Fig.1.
 
 <p align="center">
-<img src="../images/ds2onlineModel.png" width=800>
+<img src="../../images/ds2onlineModel.png" width=800>
 <br/>Fig.1 The Arcitecture of deepspeech2 online model
 </p>
 
+
 ### Data Preparation
 #### Vocabulary
 For English data, the vocabulary dictionary is composed of 26 English characters with " ' ", space, \<blank\> and \<eos\>. The \<blank\> represents the blank label in CTC, the \<unk\> represents the unknown character and the \<eos\> represents the start and the end characters. For mandarin, the vocabulary dictionary is composed of chinese characters statisticed from the training set and three additional characters are added. The added characters are \<blank\>, \<unk\> and \<eos\>.  For both English and mandarin data, we set the default indexs that \<blank\>=0, \<unk\>=1 and \<eos\>= last index.
@@ -130,7 +131,7 @@ By using the command above, the training process can be started. There are 5 sta
 Using the command below, you can test the deepspeech2 online model.
  ```
  bash run.sh --stage 3 --stop_stage 5 --model_type online --conf_path conf/deepspeech2_online.yaml
-```
+ ```
 The detail commands are:
 ```
 conf_path=conf/deepspeech2_online.yaml
@@ -152,7 +153,7 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # test export ckpt avg_n
     CUDA_VISIBLE_DEVICES=0 ./local/test_export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}|| exit -1
 fi
- ```
+```
 After the training process, we use stage 3,4,5 for testing process. The stage 3 is for testing the model generated in the stage 2 and provided the CER index of the test set. The stage 4 is for transforming the model from dynamic graph to static graph by using "paddle.jit" library. The stage 5 is for testing the model in static graph.
 
 
@@ -161,12 +162,13 @@ The deepspeech2 offline model is similarity to the deepspeech2 online model. The
 
 The arcitecture of the model is shown in Fig.2.
 <p align="center">
-<img src="../images/ds2offlineModel.png" width=800>
+<img src="../../images/ds2offlineModel.png" width=800>
 <br/>Fig.2 The Arcitecture of deepspeech2 offline model
 </p>
 
 
 
+
 For data preparation and decoder, the deepspeech2 offline model is same with the deepspeech2 online model.
 
 The code of encoder and decoder for deepspeech2 offline model is in:
@@ -182,7 +184,7 @@ For training and testing, the "model_type" and the "conf_path" must be set.
 # Training offline
 cd examples/aishell/s0
 bash run.sh --stage 0 --stop_stage 2 --model_type offline --conf_path conf/deepspeech2.yaml
-```
+ ```
 ```
 # Testing offline
 cd examples/aishell/s0

docs/source/tts/README.md

@@ -2,10 +2,11 @@
 Parakeet aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It is built on PaddlePaddle dynamic graph and includes many influential TTS models.  
 
 <div align="center">
-  <img src="docs/images/logo.png" width=300 /> <br>
+  <img src="../../images/logo.png" width=300 /> <br>
 </div>
 
-## News  <img src="./docs/images/news_icon.png" width="40"/>
+
+## News  <img src="../../images/news_icon.png" width="40"/>
 - Oct-12-2021, Refector examples code.
 - Oct-12-2021, Parallel WaveGAN with LJSpeech. Check [examples/GANVocoder/parallelwave_gan/ljspeech](./examples/GANVocoder/parallelwave_gan/ljspeech).
 - Oct-12-2021, FastSpeech2/FastPitch with LJSpeech. Check [examples/fastspeech2/ljspeech](./examples/fastspeech2/ljspeech).

examples/aishell/README.md

@@ -5,7 +5,8 @@
 
 ## Data
 
-| Data Subset | Duration in Seconds |
-| data/manifest.train |  1.23 ~ 14.53125 |
-| data/manifest.dev  | 1.645 ~ 12.533 |  
-| data/manifest.test | 1.859125 ~ 14.6999375 |
+| Data Subset         | Duration in Seconds   |
+| ------------------- | --------------------- |
+| data/manifest.train | 1.23 ~ 14.53125       |
+| data/manifest.dev   | 1.645 ~ 12.533        |
+| data/manifest.test  | 1.859125 ~ 14.6999375 |

examples/aishell/s0/conf/deepspeech2.yaml

@@ -41,7 +41,7 @@ model:
   use_gru: True 
   share_rnn_weights: False
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
 
 training:
   n_epoch: 80

examples/aishell/s0/conf/deepspeech2_online.yaml

@@ -43,7 +43,7 @@ model:
   fc_layers_size_list: -1,
   use_gru: False 
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
   
 training:
   n_epoch: 50

examples/aishell/s1/conf/chunk_conformer.yaml

@@ -77,7 +77,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/aishell/s1/conf/conformer.yaml

@@ -72,7 +72,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/aishell/s1/local/test_hub.sh

+#!/bin/bash
+
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+config_path=$1
+ckpt_prefix=$2
+audio_file=$3
+
+chunk_mode=false
+if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
+    chunk_mode=true
+fi
+
+# download language model
+#bash local/download_lm_ch.sh
+#if [ $? -ne 0 ]; then
+#    exit 1
+#fi
+
+
+
+for type in  attention_rescoring; do
+    echo "decoding ${type}"
+    batch_size=1
+    output_dir=${ckpt_prefix}
+    mkdir -p ${output_dir}
+    python3 -u ${BIN_DIR}/test_hub.py \
+    --nproc ${ngpu} \
+    --config ${config_path} \
+    --result_file ${output_dir}/${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} \
+    --opts decoding.batch_size ${batch_size} \
+    --audio_file ${audio_file}
+
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
+exit 0

examples/aishell/s1/run.sh

@@ -13,6 +13,8 @@ avg_ckpt=avg_${avg_num}
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
 echo "checkpoint name ${ckpt}"
 
+audio_file="data/tmp.wav"
+
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     bash ./local/data.sh || exit -1
@@ -48,3 +50,8 @@ fi
     # train lm and build TLG
     ./local/tlg.sh --corpus aishell --lmtype srilm
  fi
+
+if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+    # test a single .wav file
+    CUDA_VISIBLE_DEVICES=3 ./local/test_hub.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+fi

examples/aishell3/tts3/README.md

@@ -218,6 +218,8 @@ fastspeech2_nosil_aishell3_ckpt_0.4
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained fastspeech2 and parallel wavegan models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/multi_spk_synthesize_e2e.py \

examples/aishell3/vc0/README.md

@@ -2,7 +2,7 @@
 This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of  [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf) . The general steps are as follows:
 1. Speaker Encoder: We  use a Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2, because the  transcriptions are not needed, we use more datasets, refer to  [ge2e](../../other/ge2e).
 2. Synthesizer: Then, we use the trained speaker encoder to generate utterance embedding for each  sentence in AISHELL-3. This embedding is a extra input of  Tacotron2 which will be concated with encoder outputs.
-3. Vocoder: We use WaveFlow as the neural Vocoder，参考实验 [waveflow](../../ljspeech/voc0).
+3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](../../ljspeech/voc0).
 
 ## Get Started
 Assume the path to the dataset is `~/datasets/data_aishell3`.

examples/aishell3/vc0/local/preprocess.sh

@@ -35,9 +35,3 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         --input=${preprocess_path}/normalized_wav \
         --output=${preprocess_path}/mel
 fi
-
-
-
-
-
-

examples/csmsc/tts2/README.md

@@ -223,6 +223,8 @@ speedyspeech_nosil_baker_ckpt_0.5
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained speedyspeech and parallel wavegan models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/synthesize_e2e.py \

examples/csmsc/tts3/README.md

@@ -202,6 +202,8 @@ fastspeech2_nosil_baker_ckpt_0.4
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained fastspeech2 and parallel wavegan models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/synthesize_e2e.py \

examples/librispeech/README.md

@@ -2,8 +2,7 @@
 
 * s0 is for deepspeech2 offline
 * s1 is for transformer/conformer/U2
-* s2 is for transformer/conformer/U2 w/ kaldi feat
-need install Kaldi
+* s2 is for transformer/conformer/U2 w/ kaldi feat, need install Kaldi
 
 ## Data
 | Data Subset | Duration in Seconds |

examples/librispeech/s0/README.md

 # LibriSpeech
 
-## Data
-| Data Subset | Duration in Seconds |
-| --- | --- |
-| data/manifest.train |  0.83s ~ 29.735s |
-| data/manifest.dev | 1.065 ~ 35.155s |  
-| data/manifest.test-clean | 1.285s ~ 34.955s |
-
 ## Deepspeech2
-
 | Model | Params | release |  Config | Test set | Loss | WER |  
 | --- | --- | --- | --- | --- | --- | --- |  
 | DeepSpeech2 | 42.96M | 2.2.0 | conf/deepspeech2.yaml + spec_aug | test-clean | 14.49190807 | 0.067283 |  

examples/librispeech/s0/conf/deepspeech2.yaml

@@ -41,7 +41,7 @@ model:
   use_gru: False 
   share_rnn_weights: True
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
 
 training:
   n_epoch: 50

examples/librispeech/s0/conf/deepspeech2_online.yaml

@@ -43,7 +43,7 @@ model:
   fc_layers_size_list: 512, 256
   use_gru: False 
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
 
 training:
   n_epoch: 50

examples/librispeech/s0/local/test_hub.sh

+#!/bin/bash
+
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+config_path=$1
+ckpt_prefix=$2
+model_type=$3
+audio_file=$4
+
+# download language model
+bash local/download_lm_en.sh
+if [ $? -ne 0 ]; then
+   exit 1
+fi
+
+python3 -u ${BIN_DIR}/test_hub.py \
+--nproc ${ngpu} \
+--config ${config_path} \
+--result_file ${ckpt_prefix}.rsl \
+--checkpoint_path ${ckpt_prefix} \
+--model_type ${model_type} \
+--audio_file ${audio_file}
+
+if [ $? -ne 0 ]; then
+    echo "Failed in evaluation!"
+    exit 1
+fi
+
+
+exit 0

examples/librispeech/s0/run.sh

@@ -13,6 +13,7 @@ avg_ckpt=avg_${avg_num}
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
 echo "checkpoint name ${ckpt}"
 
+audio_file="data/tmp.flac"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     bash ./local/data.sh || exit -1
@@ -37,3 +38,8 @@ if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # export ckpt avg_n
     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}
 fi
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+    # test a single .wav file
+    CUDA_VISIBLE_DEVICES=3 ./local/test_hub.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
+fi

examples/librispeech/s1/conf/chunk_conformer.yaml

@@ -77,7 +77,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s1/conf/chunk_transformer.yaml

@@ -70,7 +70,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s1/conf/conformer.yaml

@@ -73,7 +73,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s1/conf/transformer.yaml

@@ -68,7 +68,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null 
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s1/local/test_hub.sh

+#!/bin/bash
+
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+config_path=$1
+ckpt_prefix=$2
+audio_file=$3
+
+# bpemode (unigram or bpe)
+nbpe=5000
+bpemode=unigram
+bpeprefix="data/bpe_${bpemode}_${nbpe}"
+bpemodel=${bpeprefix}.model
+
+chunk_mode=false
+if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
+    chunk_mode=true
+fi
+
+# download language model
+#bash local/download_lm_ch.sh
+#if [ $? -ne 0 ]; then
+#    exit 1
+#fi
+
+
+for type in attention_rescoring; do
+    echo "decoding ${type}"
+    batch_size=1
+    output_dir=${ckpt_prefix}
+    mkdir -p ${output_dir}
+    python3 -u ${BIN_DIR}/test_hub.py \
+    --nproc ${ngpu} \
+    --config ${config_path} \
+    --result_file ${output_dir}/${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} \
+    --opts decoding.batch_size ${batch_size} \
+    --audio_file ${audio_file}
+
+    #score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel}.model --wer true ${expdir}/${decode_dir} ${dict}
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
+
+exit 0

examples/librispeech/s1/run.sh

@@ -15,6 +15,8 @@ avg_ckpt=avg_${avg_num}
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
 echo "checkpoint name ${ckpt}"
 
+audio_file="data/tmp.flac"
+
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     bash ./local/data.sh || exit -1
@@ -44,3 +46,8 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # export ckpt avg_n
     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi
+
+if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+    # test a single .wav file
+    CUDA_VISIBLE_DEVICES=3 ./local/test_hub.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+fi

examples/librispeech/s2/README.md

 # LibriSpeech
 
-## Transformer
-| Model | Params | Config | Augmentation| Test Set | Decode Method | Loss | WER % |  
-| --- | --- | --- | --- | --- | --- | --- | --- |
-| transformer | 32.52 M | conf/transformer.yaml | spec_aug | test-clean | attention | 6.395054340362549 | 4.2 |  
-| transformer | 32.52 M | conf/transformer.yaml | spec_aug | test-clean | ctc_greedy_search | 6.395054340362549 | 5.0 |  
-| transformer | 32.52 M | conf/transformer.yaml | spec_aug | test-clean | ctc_prefix_beam_search | 6.395054340362549 |  |  
-| transformer | 32.52 M | conf/transformer.yaml | spec_aug | test-clean | attention_rescore | 6.395054340362549 |  |  
+| Model | Params | Config | Augmentation| Loss |
+| --- | --- | --- | --- |
+| transformer | 32.52 M | conf/transformer.yaml | spec_aug | 6.3197922706604 |
+
+
+| Test Set | Decode Method | #Snt | #Wrd | Corr | Sub | Del | Ins | Err | S.Err |  
+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| test-clean | attention | 2620 | 52576 | 96.4 | 2.5 | 1.1 | 0.4 | 4.0 | 34.7 |  
+| test-clean | ctc_greedy_search | 2620 | 52576 | 95.9 | 3.7 | 0.4 | 0.5 | 4.6 | 48.0 |  
+| test-clean | ctc_prefix_beamsearch | 2620 | 52576 | 95.9 | 3.7 | 0.4 | 0.5 | 4.6 | 47.6 |  
+| test-clean | attention_rescore | 2620 | 52576 | 96.8 | 2.9 | 0.3 | 0.4 | 3.7 | 38.0 |  
+| test-clean | join_ctc_w/o_lm | 2620 | 52576 | 97.2 | 2.6 | 0.3 | 0.4 | 3.2 | 34.9 |  

examples/librispeech/s2/conf/chunk_conformer.yaml

@@ -77,7 +77,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s2/conf/chunk_transformer.yaml

@@ -70,7 +70,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s2/conf/conformer.yaml

@@ -73,7 +73,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/librispeech/s2/conf/decode/decode.yaml

+batchsize: 0
+beam-size: 60
+ctc-weight: 0.0
+lm-weight: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+penalty: 0.0

examples/librispeech/s2/conf/decode/decode_all.yaml

+batchsize: 0
+beam-size: 60
+ctc-weight: 0.4
+lm-weight: 0.6
+maxlenratio: 0.0
+minlenratio: 0.0
+penalty: 0.0
\ No newline at end of file

examples/librispeech/s2/conf/decode/decode_wo_lm.yaml

+batchsize: 0
+beam-size: 60
+ctc-weight: 0.4
+lm-weight: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+penalty: 0.0
\ No newline at end of file

examples/librispeech/s2/local/recog.sh

+#!/bin/bash
+
+set -e
+
+expdir=exp
+datadir=data
+nj=32
+tag=
+
+# decode config
+decode_config=conf/decode/decode.yaml
+
+# lm params
+lang_model=rnnlm.model.best
+lmexpdir=exp/train_rnnlm_pytorch_lm_transformer_cosine_batchsize32_lr1e-4_layer16_unigram5000_ngpu4/
+lmtag='nolm'
+
+recog_set="test-clean test-other dev-clean dev-other"
+recog_set="test-clean"
+
+# bpemode (unigram or bpe)
+nbpe=5000
+bpemode=unigram
+bpeprefix="data/bpe_${bpemode}_${nbpe}"
+bpemodel=${bpeprefix}.model
+
+# bin params
+config_path=conf/transformer.yaml
+dict=data/bpe_unigram_5000_units.txt
+ckpt_prefix=
+
+source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
+
+if [ -z ${ckpt_prefix} ]; then
+    echo "usage: $0 --ckpt_prefix ckpt_prefix"
+    exit 1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+ckpt_dir=$(dirname `dirname ${ckpt_prefix}`)
+echo "ckpt dir: ${ckpt_dir}"
+
+ckpt_tag=$(basename ${ckpt_prefix})
+echo "ckpt tag: ${ckpt_tag}"
+
+chunk_mode=false
+if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
+    chunk_mode=true
+fi
+echo "chunk mode: ${chunk_mode}"
+echo "decode conf: ${decode_config}"
+
+# download language model
+#bash local/download_lm_en.sh
+#if [ $? -ne 0 ]; then
+#    exit 1
+#fi
+
+
+pids=() # initialize pids
+
+for dmethd in join_ctc; do
+(
+    echo "${dmethd} decoding"
+    for rtask in ${recog_set}; do
+    (
+        echo "${rtask} dataset"
+        decode_dir=${ckpt_dir}/decode/decode_${rtask/-/_}_${dmethd}_$(basename ${config_path%.*})_${lmtag}_${ckpt_tag}_${tag}
+        feat_recog_dir=${datadir}
+        mkdir -p ${decode_dir}
+        mkdir -p ${feat_recog_dir}
+
+        # split data
+        split_json.sh manifest.${rtask} ${nj}
+
+        #### use CPU for decoding
+        ngpu=0
+
+        # set batchsize 0 to disable batch decoding
+        ${decode_cmd} JOB=1:${nj} ${decode_dir}/log/decode.JOB.log \
+            python3 -u ${BIN_DIR}/recog.py \
+                --api v2 \
+                --config ${decode_config} \
+                --ngpu ${ngpu} \
+                --batchsize 0 \
+                --checkpoint_path ${ckpt_prefix} \
+                --dict-path ${dict} \
+                --recog-json ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask} \
+                --result-label ${decode_dir}/data.JOB.json \
+                --model-conf ${config_path} \
+                --model ${ckpt_prefix}.pdparams
+
+                #--rnnlm ${lmexpdir}/${lang_model} \
+
+        score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${decode_dir} ${dict}
+
+    ) &
+    pids+=($!) # store background pids
+    i=0; for pid in "${pids[@]}"; do wait ${pid} || ((++i)); done
+    [ ${i} -gt 0 ] && echo "$0: ${i} background jobs are failed." || true
+    done
+)
+done
+
+echo "Finished"
+
+exit 0

examples/librispeech/s2/local/test.sh

@@ -6,7 +6,7 @@ expdir=exp
 datadir=data
 nj=32
 
-lmtag=
+lmtag='nolm'
 
 recog_set="test-clean test-other dev-clean dev-other"
 recog_set="test-clean"
@@ -17,23 +17,31 @@ bpemode=unigram
 bpeprefix="data/bpe_${bpemode}_${nbpe}"
 bpemodel=${bpeprefix}.model
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path dict_path ckpt_path_prefix"
-    exit -1
+config_path=conf/transformer.yaml
+dict=data/bpe_unigram_5000_units.txt
+ckpt_prefix=
+
+source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
+
+if [ -z ${ckpt_prefix} ]; then
+    echo "usage: $0 --ckpt_prefix ckpt_prefix"
+    exit 1
 fi
 
 ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
-config_path=$1
-dict=$2
-ckpt_prefix=$3
+ckpt_dir=$(dirname `dirname ${ckpt_prefix}`)
+echo "ckpt dir: ${ckpt_dir}"
+
+ckpt_tag=$(basename ${ckpt_prefix})
+echo "ckpt tag: ${ckpt_tag}"
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
     chunk_mode=true
 fi
-echo "chunk mode ${chunk_mode}"
+echo "chunk mode: ${chunk_mode}"
 
 
 # download language model
@@ -46,13 +54,13 @@ pids=() # initialize pids
 
 for dmethd in attention ctc_greedy_search ctc_prefix_beam_search attention_rescoring; do
 (
-    echo "${dmethd} decoding"
+    echo "decode method: ${dmethd}"
     for rtask in ${recog_set}; do
     (
-        echo "${rtask} dataset"
-        decode_dir=decode_${rtask/-/_}_${dmethd}_$(basename ${config_path%.*})_${lmtag}
+        echo "dataset: ${rtask}"
+        decode_dir=${ckpt_dir}/decode/decode_${rtask/-/_}_${dmethd}_$(basename ${config_path%.*})_${lmtag}_${ckpt_tag}
         feat_recog_dir=${datadir}
-        mkdir -p ${expdir}/${decode_dir}
+        mkdir -p ${decode_dir}
         mkdir -p ${feat_recog_dir}
 
         # split data
@@ -63,7 +71,7 @@ for dmethd in attention ctc_greedy_search ctc_prefix_beam_search attention_resco
 
         # set batchsize 0 to disable batch decoding
         batch_size=1
-        ${decode_cmd} JOB=1:${nj} ${expdir}/${decode_dir}/log/decode.JOB.log \
+        ${decode_cmd} JOB=1:${nj} ${decode_dir}/log/decode.JOB.log \
             python3 -u ${BIN_DIR}/test.py \
             --model-name u2_kaldi \
             --run-mode test \
@@ -71,12 +79,12 @@ for dmethd in attention ctc_greedy_search ctc_prefix_beam_search attention_resco
             --dict-path ${dict} \
             --config ${config_path} \
             --checkpoint_path ${ckpt_prefix} \
-            --result-file ${expdir}/${decode_dir}/data.JOB.json \
+            --result-file ${decode_dir}/data.JOB.json \
             --opts decoding.decoding_method ${dmethd} \
             --opts decoding.batch_size ${batch_size} \
             --opts data.test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask}
 
-        score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${expdir}/${decode_dir} ${dict}
+        score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${decode_dir} ${dict}
 
     ) &
     pids+=($!) # store background pids

examples/librispeech/s2/run.sh

 #!/bin/bash
+
 set -e
 
 . ./path.sh || exit 1;
@@ -7,8 +8,9 @@ set -e
 stage=0
 stop_stage=100
 conf_path=conf/transformer.yaml
-dict_path=data/train_960_unigram5000_units.txt
+dict_path=data/bpe_unigram_5000_units.txt
 avg_num=10
+
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
 avg_ckpt=avg_${avg_num}

examples/ljspeech/tts1/README.md

@@ -188,6 +188,8 @@ transformer_tts_ljspeech_ckpt_0.4
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences_en.txt` using pretrained transformer_tts  and waveflow models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/synthesize_e2e.py \

examples/ljspeech/tts3/README.md

@@ -202,6 +202,8 @@ fastspeech2_nosil_ljspeech_ckpt_0.5
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences_en.txt` using pretrained fastspeech2 and parallel wavegan models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/synthesize_e2e_en.py \

examples/other/punctuation_restoration/README.md

 # Punctation Restoration
 
-Please using [PaddleSpeechTask](https://github.com/745165806/PaddleSpeechTask] to do this task.
+Please using [PaddleSpeechTask](https://github.com/745165806/PaddleSpeechTask) to do this task.

examples/ted_en_zh/t0/conf/transformer.yaml

@@ -69,7 +69,7 @@ model:
         asr_weight: 0.0
         ctc_weight: 0.0
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/ted_en_zh/t0/conf/transformer_joint_noam.yaml

@@ -69,7 +69,7 @@ model:
         asr_weight: 0.5
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/timit/s1/conf/transformer.yaml

@@ -67,7 +67,7 @@ model:
     model_conf:
         ctc_weight: 0.5
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/tiny/s0/conf/deepspeech2.yaml

@@ -42,7 +42,7 @@ model:
   use_gru: False 
   share_rnn_weights: True 
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
 
 training:
   n_epoch: 10

examples/tiny/s0/conf/deepspeech2_online.yaml

@@ -44,7 +44,7 @@ model:
   fc_layers_size_list: 512, 256
   use_gru: True 
   blank_id: 0
-  ctc_grad_norm_type: instance
+  ctc_grad_norm_type: null
 
 training:
   n_epoch: 10

examples/tiny/s1/conf/chunk_confermer.yaml

@@ -77,7 +77,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/tiny/s1/conf/chunk_transformer.yaml

@@ -70,7 +70,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/tiny/s1/conf/conformer.yaml

@@ -73,7 +73,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/tiny/s1/conf/transformer.yaml

@@ -67,7 +67,7 @@ model:
     model_conf:
         ctc_weight: 0.3
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: null
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 

examples/vctk/tts3/README.md

@@ -212,6 +212,8 @@ fastspeech2_nosil_vctk_ckpt_0.5
 ```
 You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained fastspeech2 and parallel wavegan models.
 ```bash
+source path.sh
+
 FLAGS_allocator_strategy=naive_best_fit \
 FLAGS_fraction_of_gpu_memory_to_use=0.01 \
 python3 ${BIN_DIR}/multi_spk_synthesize_e2e_en.py \

requirements.txt

+ConfigArgParse
 coverage
 editdistance
 g2p_en

tools/Makefile

@@ -6,7 +6,7 @@ CC ?= gcc        # used for sph2pipe
 # CXX = clang++  # Uncomment these lines...
 # CC = clang     # ...to build with Clang.
 
-$(WGET) ?= wget --no-check-certificate
+WGET ?= wget --no-check-certificate
 
 .PHONY: all clean
 
@@ -85,7 +85,7 @@ sctk-$(SCTK_GITHASH).tar.gz:
 	if [ -d '$(DOWNLOAD_DIR)' ]; then \
 	  cp -p '$(DOWNLOAD_DIR)/sctk-$(SCTK_GITHASH).tar.gz' .; \
 	else \
-	  $($(WGET)) -nv -T 10 -t 3 -O sctk-$(SCTK_GITHASH).tar.gz \
+	  $(WGET) -nv -T 10 -t 3 -O sctk-$(SCTK_GITHASH).tar.gz \
 	    https://github.com/usnistgov/SCTK/archive/$(SCTK_GITHASH).tar.gz; \
 	fi
 

utils/avg_model.py

@@ -25,8 +25,8 @@ def main(args):
     paddle.set_device('cpu')
 
     val_scores = []
-    beat_val_scores = []
-    selected_epochs = []
+    beat_val_scores = None
+    selected_epochs = None
 
     jsons = glob.glob(f'{args.ckpt_dir}/[!train]*.json')
     jsons = sorted(jsons, key=os.path.getmtime, reverse=True)
@@ -80,9 +80,10 @@ def main(args):
         data = json.dumps({
             "mode": 'val_best' if args.val_best else 'latest',
             "avg_ckpt": args.dst_model,
-            "ckpt": path_list,
-            "epoch": selected_epochs.tolist(),
-            "val_loss": beat_val_scores.tolist(),
+            "val_loss_mean": np.mean(beat_val_scores),
+            "ckpts": path_list,
+            "epochs": selected_epochs.tolist(),
+            "val_losses": beat_val_scores.tolist(),
         })
         f.write(data + "\n")