fix the bug: read space as unk

deepspeech/frontend/featurizer/text_featurizer.py

@@ -83,6 +83,8 @@ class TextFeaturizer():
         tokens = self.tokenize(text)
         ids = []
         for token in tokens:
+            if '' in self.vocab_dict and token == ' ':
+                token = ''
             token = token if token in self.vocab_dict else self.unk
             ids.append(self.vocab_dict[token])
         return ids

examples/v18_to_v2x/exp_aishell/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/v18_to_v2x/exp_aishell/local/data.sh

 #!/bin/bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
 
 stage=-1
 stop_stage=100
@@ -9,17 +15,16 @@ mkdir -p data
 TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf aishell_model_v1.8_to_v2.x.tar.gz
-mv aishell_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm aishell_model_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
 
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
@@ -39,27 +44,6 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 fi
 
 
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --sample_rate=16000 \
-    --use_dB_normalization=True \
-    --num_samples=2000 \
-    --num_workers=${num_workers} \
-    --output_path="data/mean_std.json"
-
-    if [ $? -ne 0 ]; then
-        echo "Compute mean and stddev failed. Terminated."
-        exit 1
-    fi
-fi
-
 
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # format manifest with tokenids, vocab size
@@ -67,7 +51,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     {
         python3 ${MAIN_ROOT}/utils/format_data.py \
                 --feat_type "raw" \
-                --cmvn_path "data/mean_std.json" \
+                --cmvn_path "data/mean_std.npz" \
                 --unit_type "char" \
                 --vocab_path="data/vocab.txt" \
                 --manifest_path="data/manifest.${dataset}.raw" \

examples/v18_to_v2x/exp_aishell/local/download_model.sh

 #! /usr/bin/env bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz'
 MD5=4ade113c69ea291b8ce5ec6a03296659
-TARGET=./aishell_model_v1.8_to_v2.x.tar.gz
+TARGET=${ckpt_dir}/aishell_model_v1.8_to_v2.x.tar.gz
 
 
 echo "Download Aishell model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download Aishell model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/v18_to_v2x/exp_aishell/path.sh

@@ -12,5 +12,4 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
-echo "BIN_DIR "${BIN_DIR}
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin

examples/v18_to_v2x/exp_aishell/run.sh

@@ -17,11 +17,11 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=1 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=2 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi
 

examples/v18_to_v2x/exp_baidu_en8k/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/v18_to_v2x/exp_baidu_en8k/local/data.sh

 #!/bin/bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
 
 stage=-1
 stop_stage=100
+unit_type=char
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -10,17 +17,17 @@ TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
 
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf baidu_en8k_v1.8_to_v2.x.tar.gz
-mv baidu_en8k_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm baidu_en8k_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
+
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     # download data, generate manifests
@@ -53,35 +60,13 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 fi
 
 
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --num_samples=2000 \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --sample_rate=16000 \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --use_dB_normalization=True \
-    --num_workers=${num_workers} \
-    --output_path="data/mean_std.json"
-
-    if [ $? -ne 0 ]; then
-        echo "Compute mean and stddev failed. Terminated."
-        exit 1
-    fi
-fi
-
-
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # format manifest with tokenids, vocab size
     for set in train dev test dev-clean dev-other test-clean test-other; do
     {
         python3 ${MAIN_ROOT}/utils/format_data.py \
         --feat_type "raw" \
-        --cmvn_path "data/mean_std.json" \
+        --cmvn_path "data/mean_std.npz" \
         --unit_type ${unit_type} \
         --vocab_path="data/vocab.txt" \
         --manifest_path="data/manifest.${set}.raw" \

examples/v18_to_v2x/exp_baidu_en8k/local/download_model.sh

 #! /usr/bin/env bash
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz'
 MD5=fdabeb6c96963ac85d9188f0275c6a1b
-TARGET=./baidu_en8k_v1.8_to_v2.x.tar.gz
+TARGET=${ckpt_dir}/baidu_en8k_v1.8_to_v2.x.tar.gz
 
 
 echo "Download BaiduEn8k model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download BaiduEn8k model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/v18_to_v2x/exp_baidu_en8k/path.sh

@@ -12,5 +12,5 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin
 echo "BIN_DIR "${BIN_DIR}

examples/v18_to_v2x/exp_baidu_en8k/run.sh

@@ -17,7 +17,7 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then

examples/v18_to_v2x/exp_librispeech/.gitignore

@@ -2,3 +2,4 @@ exp
 data
 *log
 tmp
+nohup*

examples/v18_to_v2x/exp_librispeech/local/data.sh

 #!/bin/bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 stage=-1
 stop_stage=100
+unit_type=char
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -9,18 +17,16 @@ mkdir -p data
 TARGET_DIR=${MAIN_ROOT}/examples/dataset
 mkdir -p ${TARGET_DIR}
 
-
-bash local/download_model.sh
+bash local/download_model.sh ${ckpt_dir}
 if [ $? -ne 0 ]; then
    exit 1
 fi
 
+cd ${ckpt_dir}
 tar xzvf librispeech_v1.8_to_v2.x.tar.gz
-mv librispeech_v1.8.pdparams exp/deepspeech2/checkpoints/
-mv README.md exp/deepspeech2/
-mv mean_std.npz data/
-mv vocab.txt data/
-rm librispeech_v1.8_to_v2.x.tar.gz -f
+cd -
+mv ${ckpt_dir}/mean_std.npz data/
+mv ${ckpt_dir}/vocab.txt data/
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     # download data, generate manifests
@@ -52,36 +58,13 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     done
 fi
 
-
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --num_samples=2000 \
-    --specgram_type="linear" \
-    --delta_delta=false \
-    --sample_rate=16000 \
-    --stride_ms=10.0 \
-    --window_ms=20.0 \
-    --use_dB_normalization=True \
-    --num_workers=${num_workers} \
-    --output_path="data/mean_std.json"
-
-    if [ $? -ne 0 ]; then
-        echo "Compute mean and stddev failed. Terminated."
-        exit 1
-    fi
-fi
-
-
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # format manifest with tokenids, vocab size
     for set in train dev test dev-clean dev-other test-clean test-other; do
     {
         python3 ${MAIN_ROOT}/utils/format_data.py \
         --feat_type "raw" \
-        --cmvn_path "data/mean_std.json" \
+        --cmvn_path "data/mean_std.npz" \
         --unit_type ${unit_type} \
         --vocab_path="data/vocab.txt" \
         --manifest_path="data/manifest.${set}.raw" \

examples/v18_to_v2x/exp_librispeech/local/download_model.sh

 #! /usr/bin/env bash
 
+if [ $# != 1 ];then
+    echo "usage: ${0} ckpt_dir"
+    exit -1
+fi
+
+ckpt_dir=$1
+
 . ${MAIN_ROOT}/utils/utility.sh
 
 URL='https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz'
 MD5=7b0f582fe2f5a840b840e7ee52246bc5
-TARGET=./librispeech_v1.8_to_v2.x.tar.gz
+TARGET=${ckpt_dir}/librispeech_v1.8_to_v2.x.tar.gz
 
 
 echo "Download LibriSpeech model ..."
@@ -13,7 +20,6 @@ if [ $? -ne 0 ]; then
     echo "Fail to download LibriSpeech model!"
     exit 1
 fi
-tar -zxvf $TARGET
 
 
 exit 0

examples/v18_to_v2x/exp_librispeech/path.sh

@@ -12,5 +12,5 @@ export PYTHONPATH=${LOCAL_DEEPSPEECH2}:${PYTHONPATH}
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=deepspeech2
-export BIN_DIR=${LOCAL_DEEPSPEECH2}/deepspeech2x/bin
+export BIN_DIR=${LOCAL_DEEPSPEECH2}/src_deepspeech2x/bin
 echo "BIN_DIR "${BIN_DIR}

examples/v18_to_v2x/exp_librispeech/run.sh

@@ -17,10 +17,10 @@ echo "checkpoint name ${ckpt}"
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     # prepare data
     mkdir -p exp/${ckpt}/checkpoints
-    bash ./local/data.sh || exit -1
+    bash ./local/data.sh exp/${ckpt}/checkpoints || exit -1
 fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=1 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi

examples/v18_to_v2x/deepspeech2x/bin/test.py → examples/v18_to_v2x/src_deepspeech2x/bin/test.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Evaluation for DeepSpeech2 model."""
-from deepspeech2x.model import DeepSpeech2Tester as Tester
+from src_deepspeech2x.test_model import DeepSpeech2Tester as Tester
 
 from deepspeech.exps.deepspeech2.config import get_cfg_defaults
 from deepspeech.training.cli import default_argument_parser

examples/v18_to_v2x/deepspeech2x/models/ds2/deepspeech2.py → examples/v18_to_v2x/src_deepspeech2x/models/ds2/deepspeech2.py

@@ -15,8 +15,8 @@
 from typing import Optional
 
 import paddle
-from deepspeech2x.models.ds2.rnn import RNNStack
 from paddle import nn
+from src_deepspeech2x.models.ds2.rnn import RNNStack
 from yacs.config import CfgNode
 
 from deepspeech.models.ds2.conv import ConvStack

examples/v18_to_v2x/deepspeech2x/model.py → examples/v18_to_v2x/src_deepspeech2x/test_model.py

@@ -20,10 +20,10 @@ from typing import Optional
 
 import numpy as np
 import paddle
-from deepspeech2x.models.ds2 import DeepSpeech2InferModel
-from deepspeech2x.models.ds2 import DeepSpeech2Model
 from paddle import distributed as dist
 from paddle.io import DataLoader
+from src_deepspeech2x.models.ds2 import DeepSpeech2InferModel
+from src_deepspeech2x.models.ds2 import DeepSpeech2Model
 from yacs.config import CfgNode
 
 from deepspeech.io.collator import SpeechCollator

utils/format_data.py

@@ -53,7 +53,8 @@ def main():
     fout = open(args.output_path, 'w', encoding='utf-8')
 
     # get feat dim
-    mean, std = load_cmvn(args.cmvn_path, filetype='json')
+    filetype = args.cmvn_path.split(".")[-1]
+    mean, istd = load_cmvn(args.cmvn_path, filetype=filetype)
     feat_dim = mean.shape[0] #(D)
     print(f"Feature dim: {feat_dim}")