支持：
- 提取干净人声
- 有人声总时长
- 无人声总时长
- 最大无人声区间时长
fromsilero_vadimportload_silero_vad,get_speech_timestamps,collect_chunksfromsrc.ultisimportload_audiodefpurified_voice(self,audio_source,sample_rate=16000,min_silence_duration_ms=700,speech_pad_ms=100,output_path=None):"""声音/声纹提纯 Args: audio_source (str | Path | bytes | np.ndarray | torch.Tensor): 支持路径、字节流、Numpy、Tensor. sample_rate (int, optional): Defaults to 16000. min_silence_duration_ms (int, optional): 最少静音持续时间, Defaults to 700. speech_pad_ms (int, optional): 人声边缘缓冲, Defaults to 100. output_path (str, optional): Defaults to None. Returns: Dict: - torch.Tensor: [1, t]. - bool: True 为有人声且提纯; False 为无人声返回原音. - float: 有人声时长 (秒). - float: 无人声时长 (秒). - float: 最大无人声区间时长 (秒). """waveform,sr,total_frames,duration_time=load_audio(audio_source,target_sr=sample_rate)ori_waveform=waveform.flatten()speech_timestamps=get_speech_timestamps(ori_waveform,self.vad_model,sampling_rate=sample_rate,threshold=0.5,min_silence_duration_ms=min_silence_duration_ms,speech_pad_ms=speech_pad_ms)speech_duration=sum(((seg["end"]-seg["start"])/sample_rateforseginspeech_timestamps),0.0)silence_duration=max(duration_time-speech_duration,0.0)max_silence_interval=0.0ifnotspeech_timestamps:max_silence_interval=duration_timeelse:max_silence_interval=max(max_silence_interval,speech_timestamps[0]["start"]/sample_rate)foriinrange(len(speech_timestamps)-1):gap=(speech_timestamps[i+1]["start"]-speech_timestamps[i]["end"])/sample_rate max_silence_interval=max(max_silence_interval,gap)max_silence_interval=max(max_silence_interval,(len(ori_waveform)-speech_timestamps[-1]["end"])/sample_rate)ifspeech_timestamps:purified_waveform=collect_chunks(speech_timestamps,ori_waveform)purified_waveform=purified_waveform[None,:]ifoutput_path:torchaudio.save(output_path,purified_waveform,sample_rate,encoding="PCM_S",bits_per_sample=16)ifself.debug:print(f"提纯完成, waveform shape:{ori_waveform.shape}, purified_waveform shape:{purified_waveform.shape}, 已成功保存至：{output_path}")return{"waveform":purified_waveform,"has_speech":True,"speech_duration":round(speech_duration,4),"silence_duration":round(silence_duration,4),"max_silence_interval":round(max_silence_interval,4),}else:ifself.debug:print("未在音频中检测到有效人声。")return{"waveform":ori_waveform[None,:],"has_speech":False,"speech_duration":0.0,"silence_duration":duration_time,"max_silence_interval":duration_time}
)
