Merge pull request #671 from alibaba-damo-academy/dev_lhn

fix loading multi-channel mp3 file bug

funasr/bin/asr_inference_launch.py

@@ -867,7 +867,10 @@ def inference_paraformer_online(
             try:
                 raw_inputs = torchaudio.load(data_path_and_name_and_type[0])[0][0]
             except:
-                raw_inputs = torch.tensor(soundfile.read(data_path_and_name_and_type[0])[0])
+                raw_inputs = soundfile.read(data_path_and_name_and_type[0], dtype='float32')[0]
+                if raw_inputs.ndim == 2:
+                    raw_inputs = raw_inputs[:, 0]
+                raw_inputs = torch.tensor(raw_inputs)
         if data_path_and_name_and_type is None and raw_inputs is not None:
             if isinstance(raw_inputs, np.ndarray):
                 raw_inputs = torch.tensor(raw_inputs)

funasr/datasets/iterable_dataset.py

@@ -71,7 +71,10 @@ def load_wav(input):
     try:
         return torchaudio.load(input)[0].numpy()
     except:
-        return np.expand_dims(soundfile.read(input)[0], axis=0)
+        waveform, _ = soundfile.read(input, dtype='float32')
+        if waveform.ndim == 2:
+            waveform = waveform[:, 0]
+        return np.expand_dims(waveform, axis=0)
 
 DATA_TYPES = {
     "sound": load_wav,

funasr/datasets/large_datasets/dataset.py

@@ -128,7 +128,9 @@ class AudioDataset(IterableDataset):
                         try:
                             waveform, sampling_rate = torchaudio.load(path)
                         except:
-                            waveform, sampling_rate = soundfile.read(path)
+                            waveform, sampling_rate = soundfile.read(path, dtype='float32')
+                            if waveform.ndim == 2:
+                                waveform = waveform[:, 0]
                             waveform = np.expand_dims(waveform, axis=0)
                             waveform = torch.tensor(waveform)
                         if self.frontend_conf is not None:

funasr/utils/wav_utils.py

@@ -166,7 +166,9 @@ def compute_fbank(wav_file,
         try:
             waveform, audio_sr = torchaudio.load(wav_file)
         except:
-            waveform, audio_sr = soundfile.read(wav_file)
+            waveform, audio_sr = soundfile.read(wav_file, dtype='float32')
+            if waveform.ndim == 2:
+                waveform = waveform[:, 0]
             waveform = torch.tensor(np.expand_dims(waveform, axis=0))
         waveform = waveform * (1 << 15)
         waveform = torch_resample(waveform, audio_sr, model_sr)