[datapipe] Support wenet datapipe

examples/hi_xiaowen/s0/conf/ds_tcn.yaml

@@ -5,8 +5,8 @@ dataset_conf:
     resample_conf:
         resample_rate: 16000
     speed_perturb: false
-    feature_extraction_conf:
-        feature_type: 'fbank'
+    feats_type: 'fbank'
+    fbank_conf:
         num_mel_bins: 40
         frame_shift: 10
         frame_length: 25
@@ -20,6 +20,7 @@ dataset_conf:
     shuffle: true
     shuffle_conf:
         shuffle_size: 1500
+    sort: false
     batch_conf:
         batch_size: 256
 

examples/hi_xiaowen/s0/conf/ds_tcn_ctc.yaml

@@ -2,11 +2,15 @@ dataset_conf:
     filter_conf:
         max_length: 2048
         min_length: 0
+        token_max_length: 200
+        token_min_length: 1
+        max_output_input_ratio: 1
+        min_output_input_ratio: 0.0005
     resample_conf:
         resample_rate: 16000
     speed_perturb: false
-    feature_extraction_conf:
-        feature_type: 'fbank'
+    feats_type: 'fbank'
+    fbank_conf:
         num_mel_bins: 40
         frame_shift: 10
         frame_length: 25
@@ -20,8 +24,9 @@ dataset_conf:
     shuffle: true
     shuffle_conf:
         shuffle_size: 1500
+    sort: false
     batch_conf:
-        batch_size: 256
+        batch_size: 64
 
 model:
     hidden_dim: 256

examples/hi_xiaowen/s0/local/prepare_data.py

@@ -18,22 +18,29 @@ def main():
                         type=str,
                         help='dir containing all the wav files')
     parser.add_argument('path', type=str, help='path to the json file')
+    parser.add_argument('dict', type=str, help='path to the dict file')
     parser.add_argument('out_dir', type=str, help='out dir')
     args = parser.parse_args()
 
+    id2token = {}
+    with open(args.dict, 'r', encoding='utf-8') as f:
+        for line in f:
+            token, idx = line.strip().split()
+            id2token[int(idx)] = token
+
     with open(args.path, 'r', encoding='utf-8') as f:
         data = json.load(f)
-        utt_id, label = [], []
+        utt_id, text = [], []
         for entry in data:
             utt_id.append(entry['utt_id'])
-            label.append(int(entry['keyword_id']))
+            text.append(id2token[entry['keyword_id']])
 
     abs_dir = os.path.abspath(args.wav_dir)
     wav_path = os.path.join(args.out_dir, 'wav.scp')
     text_path = os.path.join(args.out_dir, 'text')
     with open(wav_path, 'w', encoding='utf-8') as f_wav, \
          open(text_path, 'w', encoding='utf-8') as f_text:
-        for utt, l in zip(utt_id, label):
+        for utt, l in zip(utt_id, text):
             f_wav.write('{} {}\n'.format(utt,
                                          os.path.join(abs_dir, utt + ".wav")))
             f_text.write('{} {}\n'.format(utt, l))

examples/hi_xiaowen/s0/run.sh

@@ -32,17 +32,18 @@ fi
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
   echo "Preparing datasets..."
   mkdir -p dict
-  echo "<filler> -1" > dict/words.txt
-  echo "Hi_Xiaowen 0" >> dict/words.txt
-  echo "Nihao_Wenwen 1" >> dict/words.txt
+  echo "<FILLER> -1" > dict/dict.txt
+  echo "<HI_XIAOWEN> 0" >> dict/dict.txt
+  echo "<NIHAO_WENWEN> 1" >> dict/dict.txt
+  awk '{print $1}' dict/dict.txt > dict/words.txt
 
   for folder in train dev test; do
     mkdir -p data/$folder
     for prefix in p n; do
       mkdir -p data/${prefix}_$folder
       json_path=$download_dir/mobvoi_hotword_dataset_resources/${prefix}_$folder.json
       local/prepare_data.py $download_dir/mobvoi_hotword_dataset $json_path \
-        data/${prefix}_$folder
+        dict/dict.txt data/${prefix}_$folder
     done
     cat data/p_$folder/wav.scp data/n_$folder/wav.scp > data/$folder/wav.scp
     cat data/p_$folder/text data/n_$folder/text > data/$folder/text
@@ -82,6 +83,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
       --num_keywords $num_keywords \
       --min_duration 50 \
       --seed 666 \
+      --dict ./dict \
       $cmvn_opts \
       ${checkpoint:+--checkpoint $checkpoint}
 fi
@@ -101,10 +103,11 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     --gpu 0 \
     --batch_size 256 \
     --checkpoint $score_checkpoint \
-    --score_file $result_dir/score.txt  \
+    --score_file $result_dir/score.txt \
+    --dict ./dict \
     --num_workers 8
 
-  for keyword in 0 1; do
+  for keyword in `tail -n +2 dict/words.txt`; do
     python wekws/bin/compute_det.py \
       --keyword $keyword \
       --test_data data/test/data.list \
@@ -115,8 +118,12 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
 
   # plot det curve
   python wekws/bin/plot_det_curve.py \
-      --keywords_dict dict/words.txt \
+      --keywords_dict dict/dict.txt \
       --stats_dir  $result_dir \
+      --xlim 2 \
+      --x_step 1 \
+      --ylim 5 \
+      --y_step 1 \
       --figure_file $result_dir/det.png
 fi
 

examples/hi_xiaowen/s0/run_ctc.sh

@@ -43,17 +43,18 @@ fi
 if [ ${stage} -le -2 ] && [ ${stop_stage} -ge -2 ]; then
   echo "Preparing datasets..."
   mkdir -p dict
-  echo "<filler> -1" > dict/words.txt
-  echo "Hi_Xiaowen 0" >> dict/words.txt
-  echo "Nihao_Wenwen 1" >> dict/words.txt
+  echo "<FILLER> -1" > dict/dict.txt
+  echo "<HI_XIAOWEN> 0" >> dict/dict.txt
+  echo "<NIHAO_WENWEN> 1" >> dict/dict.txt
+  awk '{print $1}' dict/dict.txt > dict/words.txt
 
   for folder in train dev test; do
     mkdir -p data/$folder
     for prefix in p n; do
       mkdir -p data/${prefix}_$folder
       json_path=$download_dir/mobvoi_hotword_dataset_resources/${prefix}_$folder.json
       local/prepare_data.py $download_dir/mobvoi_hotword_dataset $json_path \
-        data/${prefix}_$folder
+        dict/dict.txt data/${prefix}_$folder
     done
     cat data/p_$folder/wav.scp data/n_$folder/wav.scp > data/$folder/wav.scp
     cat data/p_$folder/text data/n_$folder/text > data/$folder/text
@@ -74,8 +75,8 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 
   # and we also copy the tokens and lexicon that used in
   # https://modelscope.cn/models/damo/speech_charctc_kws_phone-xiaoyun/summary
-  cp mobvoi_kws_transcription/tokens.txt data/tokens.txt
-  cp mobvoi_kws_transcription/lexicon.txt data/lexicon.txt
+  awk '{print $1, $2-1}' mobvoi_kws_transcription/tokens.txt > dict/dict.txt
+  echo '<SILENCE>' > dict/words.txt
 
 fi
 
@@ -90,9 +91,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
 
     # Here we use tokens.txt and lexicon.txt to convert txt into index
     tools/make_list.py data/$x/wav.scp data/$x/text \
-      data/$x/wav.dur data/$x/data.list  \
-      --token_file data/tokens.txt \
-      --lexicon_file data/lexicon.txt
+      data/$x/wav.dur data/$x/data.list
   done
 fi
 

requirements.txt

@@ -15,3 +15,5 @@ pyflakes==2.2.0
 lmdb
 scipy
 tqdm
+langid
+pypinyin

tools/compute_cmvn_stats.py

@@ -105,10 +105,9 @@ def __getitem__(self, idx):
 
     with open(args.train_config, 'r') as fin:
         configs = yaml.load(fin, Loader=yaml.FullLoader)
-    feat_dim = configs['dataset_conf']['feature_extraction_conf'][
+    feat_type = configs['dataset_conf']['feats_type']
+    feat_dim = configs['dataset_conf'][f'{feat_type}_conf'][
         'num_mel_bins']
-    feat_type = configs['dataset_conf']['feature_extraction_conf'][
-        'feature_type']
     resample_rate = 0
     if 'resample_conf' in configs['dataset_conf']:
         resample_rate = configs['dataset_conf']['resample_conf'][