Python语音助手静音检测异常：is_silence无语音时返回False问题

问题背景 #

正在开发小型Python语音助手，基于Vosk实现语音识别、PyAudio录制音频、Numpy计算音频能量、Wave保存音频文件。当前在实现静音检测判断用户语句结束时遇到异常：即使不说话（仅存在轻微背景噪音），is_silence方法仍返回False。尝试过更换pydub、webrtcvad等工具，以及调整阈值参数，均未解决问题。运行环境为MacOS + Python3.12虚拟环境。

原始代码片段 #

import vosk
import json
import pyaudio
import numpy as np
import wave

def is_silence(audio_chunk, silence_thresh=-50, energy_thresh=0.001):
    # Convert the audio chunk to a NumPy array of 16-bit integers
    audio_chunk = np.frombuffer(audio_chunk, dtype=np.int16)

    # Calculate the energy of the audio chunk
    energy = np.sum(audio_chunk.astype(np.float32) ** 2) / float(len(audio_chunk))

    # Check if both energy and maximum absolute value are below thresholds
    return energy < energy_thresh and np.max(np.abs(audio_chunk)) < silence_thresh

def record_audio(output_file, max_silence_duration=1500, sample_rate=16000, chunk_size=1024, silence_thresh=-50):
    # Initialize PyAudio
    p = pyaudio.PyAudio()

    # Open a stream for recording audio
    stream = p.open(format=pyaudio.paInt16,
                    channels=1,
                    rate=sample_rate,
                    input=True,
                    frames_per_buffer=chunk_size)

    # Initialize variables for storing audio frames and counting consecutive silence
    frames = []
    consecutive_silence_counter = 0

    print("Recording...")

    while True:
        # Read a chunk of audio data from the stream
        data = stream.read(chunk_size)
        frames.append(data)

        # Convert the audio chunk to a NumPy array
        audio_chunk = np.frombuffer(data, dtype=np.int16)

        # Check if the audio chunk is silence
        if is_silence(audio_chunk, silence_thresh=silence_thresh):
            consecutive_silence_counter += 1
        else:
            consecutive_silence_counter = 0

        # Break the loop if consecutive silence reaches the specified duration
        if consecutive_silence_counter * chunk_size >= (max_silence_duration // chunk_size):
            break

    print("Recording finished.")

    # Stop and close the audio stream, and terminate PyAudio
    stream.stop_stream()
    stream.close()
    p.terminate()

    # Write the recorded frames to a WAV file
    with wave.open(output_file, 'wb') as wf:
        wf.setnchannels(1)
        wf.setsampwidth(p.get_sample_size(pyaudio.paInt16))
        wf.setframerate(sample_rate)
        wf.writeframes(b''.join(frames))

# recognize_speech function

if __name__ == "__main__":
    # Path to the Vosk model
    model_path = "path/to/vosk-model-en-us-0.22-lgraph"
    
    # Path to the output audio file
    output_file_path = "path to audio file"
    
    # Record audio and save it to the specified output file
    audio_data = record_audio(output_file=output_file_path)
    
    # Recognize speech from the recorded audio using the Vosk model
    recognized_text = recognize_speech(model_path, audio_data)

    # Print the recognized text
    print("Recognized text:", recognized_text)

核心问题排查与解决方案 #

1. 修复阈值判断逻辑错误 #

原is_silence函数存在致命逻辑错误：

• 16位音频的取值范围是-32768到32767，np.max(np.abs(audio_chunk))得到的是非负的绝对值最大值，但你设置的silence_thresh=-50是负数，导致这个条件永远为False，最终整个函数返回False。

修改后的is_silence函数：

def is_silence(audio_chunk, silence_thresh=50, energy_thresh=0.001):
    audio_chunk = np.frombuffer(audio_chunk, dtype=np.int16)
    # 计算音频能量
    energy = np.sum(audio_chunk.astype(np.float32) ** 2) / float(len(audio_chunk))
    # 获取音频块的最大绝对值
    max_abs_value = np.max(np.abs(audio_chunk))
    
    # 调试输出（可选，用于观察实际数值调整阈值）
    print(f"当前能量: {energy:.6f}, 最大绝对值: {max_abs_value}")
    
    # 修正逻辑：判断绝对值最大值小于正数阈值，同时能量低于阈值
    return energy < energy_thresh and max_abs_value < silence_thresh

2. 修正录音停止条件的计算错误 #

原record_audio函数中，停止录制的条件单位不匹配：

• consecutive_silence_counter是音频块的数量，chunk_size是每个块的采样数，两者相乘无法直接和毫秒级的max_silence_duration比较。

修改后的停止判断逻辑：

def record_audio(output_file, max_silence_duration=1500, sample_rate=16000, chunk_size=1024, silence_thresh=50):
    p = pyaudio.PyAudio()
    stream = p.open(format=pyaudio.paInt16,
                    channels=1,
                    rate=sample_rate,
                    input=True,
                    frames_per_buffer=chunk_size)

    frames = []
    consecutive_silence_counter = 0
    # 计算单个音频块的时长（毫秒）
    chunk_duration_ms = (chunk_size / sample_rate) * 1000

    print("Recording...")

    while True:
        data = stream.read(chunk_size)
        frames.append(data)

        if is_silence(data, silence_thresh=silence_thresh):
            consecutive_silence_counter += 1
        else:
            consecutive_silence_counter = 0

        # 修正：连续静音总时长 >= 设置的最大静音时长时停止
        if consecutive_silence_counter * chunk_duration_ms >= max_silence_duration:
            break

    print("Recording finished.")
    stream.stop_stream()
    stream.close()
    p.terminate()

    with wave.open(output_file, 'wb') as wf:
        wf.setnchannels(1)
        wf.setsampwidth(p.get_sample_size(pyaudio.paInt16))
        wf.setframerate(sample_rate)
        wf.writeframes(b''.join(frames))

3. 适配MacOS音频特性 #

MacOS默认音频输入增益可能过高，导致背景噪音被放大。可通过两种方式优化：

• 手动调整系统音频输入增益：打开「系统设置」→「声音」→「输入」，降低麦克风增益。
• 代码中添加音频归一化处理（可选）：在is_silence函数中加入：

  # 归一化音频（避免增益过高导致数值异常）
  if np.max(np.abs(audio_chunk)) != 0:
      audio_chunk = audio_chunk / np.max(np.abs(audio_chunk))
  

4. 补充完整语音识别函数 #

原代码中recognize_speech函数未实现，补充完整逻辑：

def recognize_speech(model_path, audio_file_path):
    model = vosk.Model(model_path)
    rec = vosk.KaldiRecognizer(model, 16000)
    
    with wave.open(audio_file_path, "rb") as wf:
        while True:
            data = wf.readframes(4000)
            if len(data) == 0:
                break
            if rec.AcceptWaveform(data):
                result = json.loads(rec.Result())
                return result.get('text', '')
        # 获取最终识别结果
        final_result = json.loads(rec.FinalResult())
        return final_result.get('text', '')

完整修正后的代码 #

整合所有修改后的完整代码可直接运行测试，根据调试输出的能量和绝对值数值，微调silence_thresh和energy_thresh参数即可适配你的环境。

内容的提问来源于stack exchange，提问作者Dima Paharski