funasr2

游雁

2023-12-06 27f31cd42bb4e20dc19de0034fc0d80b449f1db1

 funasr/datasets/dataset_jsonl.py

@@ -88,18 +88,16 @@


class AudioDataset(torch.utils.data.Dataset):
   def __init__(self, path, frontend=None, tokenizer=None, token_id_converter=None):

   def __init__(self, path, frontend=None, tokenizer=None, int_pad_value: int = -1, float_pad_value: float = 0.0, **kwargs):
      super().__init__()
      self.indexed_dataset = IndexedDatasetJsonl(path)
      self.frontend = frontend.forward
      self.fs = 16000 if frontend is None else frontend.fs
      self.data_type = "sound"
      self.tokenizer = tokenizer
      self.token_id_converter = token_id_converter

      self.int_pad_value = -1
      self.float_pad_value = 0.0
      self.int_pad_value = int_pad_value
      self.float_pad_value = float_pad_value

   

@@ -115,8 +113,7 @@
      data_src = load_audio(source, fs=self.fs)
      speech, speech_lengths = extract_features(data_src, self.data_type, self.frontend)
      target = item["target"]
      text = self.tokenizer.text2tokens(target)
      ids = self.token_id_converter.tokens2ids(text)
      ids = self.tokenizer.encode(target)
      ids_lengths = len(ids)
      text, text_lengths = torch.tensor(ids, dtype=torch.int64), torch.tensor([ids_lengths], dtype=torch.int32)