transcribe/pipelines/pipe_vad.py

from .base import MetaItem, BasePipe
from ..helpers.vadprocessor import FixedVADIterator

import numpy as np
import logging

# import noisereduce as nr


class VadPipe(BasePipe):
    vac = None
    sample_rate = 16000

    def __init__(self, in_queue=None, out_queue=None) -> None:
        super().__init__(in_queue, out_queue)
        self._offset = 0 # 处理的frame size offset
        self._status = 'END'


    def reset(self):
        self._offset = 0
        self._status = 'END'

        self.vac.reset_states()

    @classmethod
    def init(cls):
        if cls.vac is None:
            cls.vac = FixedVADIterator(
                threshold=0.6,
                sampling_rate=cls.sample_rate,
                # speech_pad_ms=10
                min_silence_duration_ms = 100,
                # speech_pad_ms = 30,
                )
            cls.vac.reset_states()


    # def reduce_noise(self, data):
    #     return nr.reduce_noise(y=data, sr=self.sample_rate)

    def _process_speech_chunk(self, source_audio:np.ndarray):
        speech_dict = self.vac(source_audio, return_seconds=False)
        if speech_dict:
            relative_start_frame = None
            relative_end_frame = None
            start_frame, end_frame = speech_dict.get("start"), speech_dict.get("end")
            if start_frame:
                relative_start_frame =start_frame - self._offset
            if end_frame:
                relative_end_frame = end_frame - self._offset
            return relative_start_frame, relative_end_frame

    def process(self, in_data: MetaItem) -> MetaItem:
        if self._offset == 0:
            self.vac.reset_states()

        # silence_audio_100ms = np.zeros(int(0.1*self.sample_rate))
        source_audio = np.frombuffer(in_data.source_audio, dtype=np.float32)
        speech_data  = self._process_speech_chunk(source_audio)

        if speech_data: # 表示有音频的变化点出现
            rel_start_frame, rel_end_frame = speech_data
            if rel_start_frame is not None and rel_end_frame is None:
                self._status = "START" # 语音开始
                target_audio = source_audio[max(rel_start_frame-100, 0):]
                logging.debug("🫸 Speech start frame: {}".format(rel_start_frame))
            elif rel_start_frame is None and rel_end_frame is not None:
                self._status = "END" # 音频结束
                target_audio = source_audio[:rel_end_frame]
                logging.debug(" 🫷Speech ended, capturing audio up to frame: {}".format(rel_end_frame))
            else:
                self._status = 'END'
                target_audio = source_audio[max(rel_start_frame-100, 0):rel_end_frame]
                logging.debug(" 🔄 Speech segment captured from frame {} to frame {}".format(rel_start_frame, rel_end_frame))
                # logging.debug("❌ No valid speech segment detected, setting status to END")
        else:
            if self._status == 'START':
                target_audio = source_audio
                # logging.debug("🔊 Continuing to capture audio as speech is still ongoing")
            else: # end
                target_audio = np.array([],dtype=np.float32)
                # self._status = 'END'
                # logging.debug("❌ No speech detected, setting status to END")

        self._offset += len(source_audio)

        in_data.audio = target_audio.tobytes()
        in_data.source_audio = b''
        in_data.speech_status = self._status
        return in_data