语音质量评估STOI代码

STOI（Short-Time Objective Intelligibility）是一种用于评估语音质量的客观指标，它可以用来衡量音频的清晰度和可懂性。以下是一份计算STOI值的Python代码示例：

import soundfile as sf
import numpy as np


def stoi(clean_audio, noisy_audio, sample_rate):
    """Calculate Short-Time Objective Intelligibility (STOI)"""
    clean, _ = sf.read(clean_audio)
    noisy, _ = sf.read(noisy_audio)
    
    assert clean.shape == noisy.shape, "Both audios should have the same shape"
    
    clean = clean / np.max(np.abs(clean))
    noisy = noisy / np.max(np.abs(noisy))
    
    frame_size = int(sample_rate * 0.030)  # 30ms
    hop_size = int(sample_rate * 0.010)  # 10ms
    n_frames = int((len(clean) - frame_size) / hop_size) + 1
    
    clean_stft = np.abs(np.array([np.fft.fft(clean[i:i+frame_size], n=frame_size) for i in range(0, len(clean)-frame_size+1, hop_size)]))
    noisy_stft = np.abs(np.array([np.fft.fft(noisy[i:i+frame_size], n=frame_size) for i in range(0, len(noisy)-frame_size+1, hop_size)]))
    
    clean_stft = clean_stft[:, :int(frame_size/2)]
    noisy_stft = noisy_stft[:, :int(frame_size/2)]
    
    eps = np.finfo(float).eps
    ratio = np.divide(clean_stft, noisy_stft + eps)
    log_ratio = np.log10(ratio)
    log_ratio[np.isnan(log_ratio)] = 0
    
    mean_log_ratio = np.mean(log_ratio, axis=1)
    modified_mean_log_ratio = np.maximum(np.minimum(mean_log_ratio, 1), -1)
    
    stoi_val = np.mean(np.power(10, -0.1 * (modified_mean_log_ratio - 1)), axis=0)
    
    return stoi_val

使用方法如下：

clean_audio = "clean.wav"
noisy_audio = "noisy.wav"
sample_rate = 16000

stoi_val = stoi(clean_audio, noisy_audio, sample_rate)
print("STOI value:", stoi_val)

其中，`clean_audio`和`noisy_audio`分别是干净语音和带噪语音的文件路径，`sample_rate`是采样率。计算结果以0到1之间的值表示STOI值，值越高表示语音质量越好。