fix the bugs

examples/speech_to_text/prep_audio_data.py

@@ -185,7 +185,7 @@ class AudioDataset(Dataset):
 
         if need_waveform:
             offset = item.get('offset', False)
-            if offset:
+            if offset is not False:
                 waveform, sample_rate = torchaudio.load(audio,
                                                         frame_offset=offset,
                                                         num_frames=item["n_frames"])
@@ -272,7 +272,11 @@ def process(args):
                 waveform, sample_rate, _ = dataset.get(idx, need_waveform=True)
                 if waveform.shape[1] == 0:
                     continue
-                features = extract_fbank_features(waveform, sample_rate, Path(features_path))
+
+                try:
+                    features = extract_fbank_features(waveform, sample_rate, Path(features_path))
+                except AssertionError:
+                    logger.warning("Extract file %s failed." % utt_id)
 
                 if split == 'train' and args.cmvn_type == "global" and not utt_id.startswith("sp"):
                     if len(gcmvn_feature_list) < args.gcmvn_max_num:
@@ -326,16 +330,21 @@ def process(args):
                 _, sample_rate, n_frames = dataset.get(idx, need_waveform=False)
                 utt_id = item["id"]
 
-                manifest["id"].append(utt_id)
                 if use_raw:
                     audio_path = item["audio"]
 
                     # add offset and frames info
-                    if item.get("offset", False):
+                    if item.get("offset", False) is not False:
                         audio_path = f"{audio_path}:{item['offset']}:{n_frames}"
                     manifest["audio"].append(audio_path)
                 else:
-                    manifest["audio"].append(zip_manifest[utt_id])
+                    if utt_id in zip_manifest:
+                        manifest["audio"].append(zip_manifest[utt_id])
+                    else:
+                        logger.warning("%s is not in the zip" % utt_id)
+                        continue
+
+                manifest["id"].append(utt_id)
                 duration_ms = int(n_frames / sample_rate * 1000)
                 manifest["n_frames"].append(int(1 + (duration_ms - 25) / 10))