修复 当选用SenseVoice模型处理长音频（同时开启vad和output_timestamp）时的异常 (#2413)

hohaiuhsx

2025-03-10 6fe10a8dbfbab2bdcc28a411f9c5da85a4a8d002

修复 当选用SenseVoice模型处理长音频（同时开启vad和output_timestamp）时的异常 (#2413)

 funasr/models/sense_voice/model.py

@@ -925,11 +925,11 @@
                align = ctc_forced_align(
                    logits_speech.unsqueeze(0).float(),
                    torch.Tensor(token_int[4:]).unsqueeze(0).long().to(logits_speech.device),
                    (encoder_out_lens - 4).long(),
                    (encoder_out_lens[i] - 4).long(),
                    torch.tensor(len(token_int) - 4).unsqueeze(0).long().to(logits_speech.device),
                    ignore_id=self.ignore_id,
                )
                pred = groupby(align[0, : encoder_out_lens[0]])
                pred = groupby(align[0, : encoder_out_lens[i]])
                _start = 0
                token_id = 0
                ts_max = encoder_out_lens[i] - 4
@@ -951,20 +951,26 @@

    def post(self, timestamp):
        timestamp_new = []
        prev_word = None
        for i, t in enumerate(timestamp):
            word, start, end = t
            start = int(start * 1000)
            end = int(end * 1000)
            if word == "▁":
                continue
            if i == 0:
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start * 1000), int(end * 1000)])
            elif word.startswith("▁") or len(word) == 1 or not word[1].isalpha():
                timestamp_new.append([start, end])
            elif word.startswith("▁"):
                word = word[1:]
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start * 1000), int(end * 1000)])
                timestamp_new.append([start, end])
            elif prev_word.isalpha() and prev_word.isascii() and word.isalpha() and word.isascii():
                prev_word += word
                timestamp_new[-1][1] = end
            else:
                # timestamp_new[-1][0] += word
                timestamp_new[-1][1] = int(end * 1000)
                timestamp_new.append([start, end])
            prev_word = word
        return timestamp_new

    def export(self, **kwargs):