vad_realtime_transcribe.py

import sounddevice as sd
import webrtcvad
import numpy as np
from scipy.io.wavfile import write
from faster_whisper import WhisperModel
import time
import os

SAMPLE_RATE = 16000
FRAME_DURATION = 30  # ms
FRAME_SIZE = int(SAMPLE_RATE * FRAME_DURATION / 1000)
SILENCE_THRESHOLD = 0.5  # 秒，连续静音多久算结束
MAX_RECORD_SECONDS = 15  # 最大安全时长
MIN_SPEECH_DURATION = 0.3  # 忽略太短的无效段落

# ✅ 初始化 Whisper 模型（只加载一次）
print("📥 Loading Whisper model...")
model = WhisperModel("small", device="cpu", compute_type="int8")


def record_and_detect(filename="audio.wav"):
    vad = webrtcvad.Vad(2)
    frames = []
    silence_counter = 0
    speech_detected = False
    max_silence_frames = int(SILENCE_THRESHOLD * 1000 / FRAME_DURATION)

    stream = sd.InputStream(samplerate=SAMPLE_RATE, channels=1, dtype='int16', blocksize=FRAME_SIZE)
    stream.start()
    print("🎙️ 说话开始（说完停顿自动结束）...")

    try:
        while True:
            frame, _ = stream.read(FRAME_SIZE)
            pcm = frame.flatten()
            pcm_bytes = pcm.tobytes()
            is_speech = vad.is_speech(pcm_bytes, SAMPLE_RATE)

            frames.append((pcm.copy(), is_speech))

            if is_speech:
                silence_counter = 0
                speech_detected = True
            else:
                silence_counter += 1

            if speech_detected and silence_counter >= max_silence_frames:
                print("🛑 停顿检测完成，结束录音")
                break
    finally:
        stream.stop()
        stream.close()

    # ✅ 剪掉尾部静音帧
    cut_index = len(frames)
    for i in range(len(frames) - 1, -1, -1):
        if frames[i][1]:  # 是语音
            cut_index = i + 1
            break

    trimmed_audio = np.concatenate([frames[i][0] for i in range(cut_index)])
    duration = len(trimmed_audio) / SAMPLE_RATE

    if duration < MIN_SPEECH_DURATION:
        print("⚠️ 忽略无效短录音")
        return None

    write(filename, SAMPLE_RATE, trimmed_audio.astype(np.int16))
    print(f"💾 已保存音频：{filename} (长度: {duration:.2f}s)")
    return filename


def transcribe(filename):
    print("🔍 开始转录...")
    t1 = time.time()
    segments, info = model.transcribe(filename, beam_size=3)
    t2 = time.time()

    print(f"✅ 检测语言: {info.language}")
    segment_list = list(segments)
    if not segment_list:
        print("⚠️ 没识别到语音内容")
    else:
        print("📄 识别内容：")
        for seg in segment_list:
            print(f"[{seg.start:.2f}s → {seg.end:.2f}s] {seg.text}")
    print(f"⏱️ 转录耗时：{t2 - t1:.2f}s")


if __name__ == "__main__":
    while True:
        audio_file = record_and_detect()
        if audio_file:
            transcribe(audio_file)
        print("\n✅ 等待下一轮语音输入（Ctrl+C退出）...\n")