Merge pull request #1848 from zh794390558/spx

[speechx] rm to_float32 flags, default feature fbank

paddlespeech/cli/vector/infer.py

@@ -437,7 +437,9 @@ class VectorExecutor(BaseExecutor):
         if self.sample_rate != 16000 and self.sample_rate != 8000:
             logger.error(
                 "invalid sample rate, please input --sr 8000 or --sr 16000")
-            logger.error(f"The model sample rate: {self.sample_rate}, the external sample rate is: {sample_rate}")
+            logger.error(
+                f"The model sample rate: {self.sample_rate}, the external sample rate is: {sample_rate}"
+            )
             return False
 
         if isinstance(audio_file, (str, os.PathLike)):

paddlespeech/server/engine/vector/__init__.py

+# Copyright (c) 2022 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.

paddlespeech/server/engine/vector/python/__init__.py

+# Copyright (c) 2022 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.

paddlespeech/server/engine/vector/python/vector_engine.py

@@ -16,9 +16,9 @@ from collections import OrderedDict
 
 import numpy as np
 import paddle
-
 from paddleaudio.backends import load as load_audio
 from paddleaudio.compliance.librosa import melspectrogram
+
 from paddlespeech.cli.log import logger
 from paddlespeech.cli.vector.infer import VectorExecutor
 from paddlespeech.server.engine.base_engine import BaseEngine

speechx/examples/ds2_ol/aishell/run.sh

@@ -155,7 +155,6 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
         --wav_rspecifier=scp:$data/split${nj}/JOB/${aishell_wav_scp} \
         --cmvn_file=$cmvn \
         --model_path=$model_dir/avg_1.jit.pdmodel \
-        --to_float32=true \
         --streaming_chunk=30 \
         --param_path=$model_dir/avg_1.jit.pdiparams \
         --word_symbol_table=$wfst/words.txt \

speechx/examples/ds2_ol/decoder/recognizer_test_main.cc

@@ -19,6 +19,7 @@
 
 DEFINE_string(wav_rspecifier, "", "test feature rspecifier");
 DEFINE_string(result_wspecifier, "", "test result wspecifier");
+DEFINE_int32(sample_rate, 16000, "sample rate");
 
 int main(int argc, char* argv[]) {
     gflags::ParseCommandLineFlags(&argc, &argv, false);
@@ -30,7 +31,8 @@ int main(int argc, char* argv[]) {
     kaldi::SequentialTableReader<kaldi::WaveHolder> wav_reader(
         FLAGS_wav_rspecifier);
     kaldi::TokenWriter result_writer(FLAGS_result_wspecifier);
-    int sample_rate = 16000;
+
+    int sample_rate = FLAGS_sample_rate;
     float streaming_chunk = FLAGS_streaming_chunk;
     int chunk_sample_size = streaming_chunk * sample_rate;
     LOG(INFO) << "sr: " << sample_rate;

speechx/examples/ds2_ol/feat/compute_fbank_main.cc

@@ -69,6 +69,7 @@ int main(int argc, char* argv[]) {
     feat_cache_opts.frame_chunk_stride = 1;
     feat_cache_opts.frame_chunk_size = 1;
     ppspeech::FeatureCache feature_cache(feat_cache_opts, std::move(cmvn));
+    LOG(INFO) << "fbank: " << true;
     LOG(INFO) << "feat dim: " << feature_cache.Dim();
 
     int sample_rate = 16000;

speechx/examples/ds2_ol/feat/linear-spectrogram-wo-db-norm-ol.cc

@@ -56,6 +56,7 @@ int main(int argc, char* argv[]) {
     opt.frame_opts.remove_dc_offset = false;
     opt.frame_opts.window_type = "hanning";
     opt.frame_opts.preemph_coeff = 0.0;
+    LOG(INFO) << "linear feature: " << true;
     LOG(INFO) << "frame length (ms): " << opt.frame_opts.frame_length_ms;
     LOG(INFO) << "frame shift (ms): " << opt.frame_opts.frame_shift_ms;
 
@@ -77,7 +78,7 @@ int main(int argc, char* argv[]) {
     int sample_rate = 16000;
     float streaming_chunk = FLAGS_streaming_chunk;
     int chunk_sample_size = streaming_chunk * sample_rate;
-    LOG(INFO) << "sr: " << sample_rate;
+    LOG(INFO) << "sample rate: " << sample_rate;
     LOG(INFO) << "chunk size (s): " << streaming_chunk;
     LOG(INFO) << "chunk size (sample): " << chunk_sample_size;
 

speechx/examples/ds2_ol/websocket/websocket_server.sh

@@ -63,7 +63,6 @@ websocket_server_main \
     --cmvn_file=$cmvn \
     --model_path=$model_dir/avg_1.jit.pdmodel \
     --streaming_chunk=0.1 \
-    --to_float32=true \
     --param_path=$model_dir/avg_1.jit.pdiparams \
     --word_symbol_table=$wfst/words.txt \
     --model_output_names=softmax_0.tmp_0,tmp_5,concat_0.tmp_0,concat_1.tmp_0 \

speechx/speechx/decoder/param.h

@@ -19,23 +19,24 @@
 #include "decoder/ctc_tlg_decoder.h"
 #include "frontend/audio/feature_pipeline.h"
 
+// feature
+DEFINE_bool(use_fbank, false, "False for fbank; or linear feature");
+// DEFINE_bool(to_float32, true, "audio convert to pcm32. True for linear
+// feature, or fbank");
+DEFINE_int32(num_bins, 161, "num bins of mel");
 DEFINE_string(cmvn_file, "", "read cmvn");
 DEFINE_double(streaming_chunk, 0.1, "streaming feature chunk size");
-DEFINE_bool(to_float32, true, "audio convert to pcm32");
-DEFINE_string(model_path, "avg_1.jit.pdmodel", "paddle nnet model");
-DEFINE_string(param_path, "avg_1.jit.pdiparams", "paddle nnet model param");
-DEFINE_string(word_symbol_table, "words.txt", "word symbol table");
-DEFINE_string(graph_path, "TLG", "decoder graph");
-DEFINE_double(acoustic_scale, 1.0, "acoustic scale");
-DEFINE_int32(max_active, 7500, "max active");
-DEFINE_double(beam, 15.0, "decoder beam");
-DEFINE_double(lattice_beam, 7.5, "decoder beam");
+// feature sliding window
 DEFINE_int32(receptive_field_length,
              7,
              "receptive field of two CNN(kernel=5) downsampling module.");
 DEFINE_int32(downsampling_rate,
              4,
              "two CNN(kernel=5) module downsampling rate.");
+
+// nnet
+DEFINE_string(model_path, "avg_1.jit.pdmodel", "paddle nnet model");
+DEFINE_string(param_path, "avg_1.jit.pdiparams", "paddle nnet model param");
 DEFINE_string(
     model_input_names,
     "audio_chunk,audio_chunk_lens,chunk_state_h_box,chunk_state_c_box",
@@ -47,8 +48,14 @@ DEFINE_string(model_cache_names,
               "chunk_state_h_box,chunk_state_c_box",
               "model cache names");
 DEFINE_string(model_cache_shapes, "5-1-1024,5-1-1024", "model cache shapes");
-DEFINE_bool(use_fbank, false, "use fbank or linear feature");
-DEFINE_int32(num_bins, 161, "num bins of mel");
+
+// decoder
+DEFINE_string(word_symbol_table, "words.txt", "word symbol table");
+DEFINE_string(graph_path, "TLG", "decoder graph");
+DEFINE_double(acoustic_scale, 1.0, "acoustic scale");
+DEFINE_int32(max_active, 7500, "max active");
+DEFINE_double(beam, 15.0, "decoder beam");
+DEFINE_double(lattice_beam, 7.5, "decoder beam");
 
 namespace ppspeech {
 // todo refactor later
@@ -56,17 +63,18 @@ FeaturePipelineOptions InitFeaturePipelineOptions() {
     FeaturePipelineOptions opts;
     opts.cmvn_file = FLAGS_cmvn_file;
     opts.linear_spectrogram_opts.streaming_chunk = FLAGS_streaming_chunk;
-    opts.to_float32 = FLAGS_to_float32;
     kaldi::FrameExtractionOptions frame_opts;
     frame_opts.dither = 0.0;
     frame_opts.frame_shift_ms = 10;
     opts.use_fbank = FLAGS_use_fbank;
     if (opts.use_fbank) {
+        opts.to_float32 = false;
         frame_opts.window_type = "povey";
         frame_opts.frame_length_ms = 25;
         opts.fbank_opts.fbank_opts.mel_opts.num_bins = FLAGS_num_bins;
         opts.fbank_opts.fbank_opts.frame_opts = frame_opts;
     } else {
+        opts.to_float32 = true;
         frame_opts.remove_dc_offset = false;
         frame_opts.frame_length_ms = 20;
         frame_opts.window_type = "hanning";

speechx/speechx/frontend/audio/fbank.cc

@@ -102,13 +102,16 @@ bool Fbank::Compute(const Vector<BaseFloat>& waves, Vector<BaseFloat>* feats) {
         // note: this online feature-extraction code does not support VTLN.
         RealFft(&window, true);
         kaldi::ComputePowerSpectrum(&window);
-        const kaldi::MelBanks &mel_bank = *(computer_.GetMelBanks(1.0));
+        const kaldi::MelBanks& mel_bank = *(computer_.GetMelBanks(1.0));
         SubVector<BaseFloat> power_spectrum(window, 0, window.Dim() / 2 + 1);
         if (!opts_.fbank_opts.use_power) {
             power_spectrum.ApplyPow(0.5);
         }
-        int32 mel_offset = ((opts_.fbank_opts.use_energy && !opts_.fbank_opts.htk_compat) ? 1 : 0);
-        SubVector<BaseFloat> mel_energies(this_feature, mel_offset, opts_.fbank_opts.mel_opts.num_bins);
+        int32 mel_offset =
+            ((opts_.fbank_opts.use_energy && !opts_.fbank_opts.htk_compat) ? 1
+                                                                           : 0);
+        SubVector<BaseFloat> mel_energies(
+            this_feature, mel_offset, opts_.fbank_opts.mel_opts.num_bins);
         mel_bank.Compute(power_spectrum, &mel_energies);
         mel_energies.ApplyFloor(1e-07);
         mel_energies.ApplyLog();

speechx/speechx/frontend/audio/feature_pipeline.cc

@@ -25,11 +25,11 @@ FeaturePipeline::FeaturePipeline(const FeaturePipelineOptions& opts) {
     unique_ptr<FrontendInterface> base_feature;
 
     if (opts.use_fbank) {
-        base_feature.reset(new ppspeech::Fbank(opts.fbank_opts,
-                              std::move(data_source)));
+        base_feature.reset(
+            new ppspeech::Fbank(opts.fbank_opts, std::move(data_source)));
     } else {
-        base_feature.reset(new ppspeech::LinearSpectrogram(opts.linear_spectrogram_opts,
-                              std::move(data_source)));
+        base_feature.reset(new ppspeech::LinearSpectrogram(
+            opts.linear_spectrogram_opts, std::move(data_source)));
     }
 
     unique_ptr<FrontendInterface> cmvn(

speechx/speechx/frontend/audio/feature_pipeline.h

@@ -18,25 +18,25 @@
 
 #include "frontend/audio/audio_cache.h"
 #include "frontend/audio/data_cache.h"
+#include "frontend/audio/fbank.h"
 #include "frontend/audio/feature_cache.h"
 #include "frontend/audio/frontend_itf.h"
 #include "frontend/audio/linear_spectrogram.h"
-#include "frontend/audio/fbank.h"
 #include "frontend/audio/normalizer.h"
 
 namespace ppspeech {
 
 struct FeaturePipelineOptions {
     std::string cmvn_file;
-    bool to_float32;
+    bool to_float32;  // true, only for linear feature
     bool use_fbank;
     LinearSpectrogramOptions linear_spectrogram_opts;
     FbankOptions fbank_opts;
     FeatureCacheOptions feature_cache_opts;
     FeaturePipelineOptions()
         : cmvn_file(""),
-          to_float32(false),
-          use_fbank(false),
+          to_float32(false),  // true, only for linear feature
+          use_fbank(true),
           linear_spectrogram_opts(),
           fbank_opts(),
           feature_cache_opts() {}