hubert decode

examples/librispeech/asr3/local/test.sh

@@ -31,7 +31,7 @@ python3 utils/format_rsl.py \
 
 for type in ctc_greedy_search; do
     echo "decoding ${type}"
-    batch_size=16
+    batch_size=8
     python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \

examples/librispeech/asr3/run.sh

@@ -7,7 +7,7 @@ MODEL=hubert
 
 gpus=2
 stage=1
-stop_stage=1
+stop_stage=3
 conf_path=conf/${MODEL}ASR.yaml
 ips=            #xx.xx.xx.xx,xx.xx.xx.xx
 decode_conf_path=conf/tuning/decode.yaml
@@ -20,7 +20,7 @@ audio_file=data/demo_002_en.wav
 
 avg_ckpt=avg_${avg_num}
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
-ckpt=test3
+ckpt=test6
 echo "checkpoint name ${ckpt}"
 
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then

paddlespeech/s2t/exps/hubert/model.py

@@ -186,8 +186,8 @@ class HubertASRTrainer(Trainer):
             wavs_lens_rate = wavs_lens / wav.shape[1]
             wav = wav[:, :, 0]
 
-        # if hasattr(train_conf, 'audio_augment'):
-        #     wav = self.speech_augmentation(wav, wavs_lens_rate)
+        if hasattr(train_conf, 'audio_augment'):
+            wav = self.speech_augmentation(wav, wavs_lens_rate)
 
         loss = self.model(wav, wavs_lens_rate, target, target_lens)
 

paddlespeech/s2t/models/hubert/hubert_ASR.py

@@ -86,7 +86,6 @@ class HubertASR(nn.Layer):
         if self.normalize_wav:
             wav = F.layer_norm(wav, wav.shape)
 
-        self.hubert.eval()
         # Extract wav2vec output
         out = self.hubert.extract_features(wav)[0]
         # We normalize the output if required
@@ -205,7 +204,7 @@ class HubertASR(nn.Layer):
         if self.normalize_wav:
             wav = F.layer_norm(wav, wav.shape[1:])
         # Extract wav2vec output
-        out = self.wav2vec2(wav)[0]
+        out = self.hubert.extract_features(wav)[0]
         # We normalize the output if required
         if self.output_norm:
             out = F.layer_norm(out, out.shape[1:])
@@ -247,7 +246,7 @@ class HubertASR(nn.Layer):
         if self.normalize_wav:
             wav = F.layer_norm(wav, wav.shape[1:])
         # Extract wav2vec output
-        out = self.wav2vec2(wav)[0]
+        out = self.hubert.extract_features(wav)[0]
         # We normalize the output if required
         if self.output_norm:
             out = F.layer_norm(out, out.shape[1:])
@@ -323,14 +322,17 @@ class HubertASR(nn.Layer):
         return hyps[0][0]
 
 
-class Wav2vec2Base(nn.Layer):
+class HubertBase(nn.Layer):
     """Wav2vec2 model"""
 
     def __init__(self, config: dict):
         super().__init__()
-        wav2vec2_config = Wav2Vec2ConfigPure(config)
-        wav2vec2 = Wav2Vec2Model(wav2vec2_config)
-        self.wav2vec2 = wav2vec2
+        with open(config.vocab_filepath) as f:
+                dicts = [symbol.strip() for symbol in f.readlines()]
+        task_cfg = self.merge_with_parent(HubertPretrainingConfig, dict(self.config.task_cfg))
+        model_cfg = self.merge_with_parent(HubertConfig, dict(self.config.model_cfg))
+        hubert = HubertModel(model_cfg, task_cfg, dicts)
+        self.hubert= hubert
 
     @classmethod
     def from_config(cls, configs: dict):
@@ -340,11 +342,11 @@ class Wav2vec2Base(nn.Layer):
         Raises:
             ValueError: raise when using not support encoder type.
         Returns:
-            nn.Layer: Wav2Vec2Base
+            nn.Layer: HubertBase
         """
         model = cls(configs)
         return model
 
     def forward(self, wav):
-        out = self.wav2vec2(wav)
+        out = self.hubert(wav)
         return out