用python对一个wav文件进行logmmse算法降噪

首先，需要导入以下依赖库：

import numpy as np
import scipy.signal as signal
import scipy.io.wavfile as wavfile

接着，我们可以定义一个logmmse函数：

def logmmse(sig, rate, noise_frames=6, SSub=1, alpha=0.98, over_sub=3, noise_mu=0.1):
    # 预处理
    sig = sig.astype(np.float64)
    sig -= np.mean(sig)  # 消除DC偏置
    sig /= np.max(np.abs(sig))  # 归一化

    # 分帧
    N = 512
    frames = signal.frame(sig, frame_length=N, hop_length=N//2).astype(np.float64)
    num_frames, frame_len = frames.shape

    # 分帧FFT
    frames_fft = np.fft.fft(frames, axis=1)

    # 计算帧能量
    frame_energy = np.sum(frames ** 2, axis=1) / frame_len

    # 估计噪声能量
    noise_frames_fft = frames_fft[:noise_frames]
    noise_energy = np.sum(noise_frames_fft ** 2, axis=1) / frame_len
    noise_mu = np.median(noise_energy)
    noise_energy_db = 10 * np.log10(noise_energy)

    # 初始化变量
    alpha_s = 1
    g = np.ones(frame_len)
    xi_w = np.ones(frame_len)
    xi_w_last = 0
    sub_speech_flag = np.zeros(num_frames)
    over_sub_count = np.zeros(num_frames)

    # 主循环
    for i in range(num_frames):
        # 计算当前帧信噪比
        snr = 10 * np.log10(frame_energy[i] / noise_mu)
        snr_db = snr - 10 * np.log10(SSub)

        # 判断是否为语音帧
        if snr_db > noise_energy_db.max():
            sub_speech_flag[i] = 1

        if sub_speech_flag[i]:
            # 计算当前帧的谱增益函数
            gamma = np.minimum(np.maximum(alpha_s * (snr_db - noise_energy_db), 0), 1)

            # 调整谱增益函数
            if xi_w_last > 1:
                over_sub_count[i] = over_sub_count[i-1] + 1
            else:
                over_sub_count[i] = 0

            if over_sub_count[i] < over_sub:
                xi_w = np.power(gamma, alpha) * xi_w_last
            else:
                xi_w = np.power(gamma, alpha) * (xi_w_last + 1)
                over_sub_count[i] = over_sub

            # 计算平滑系数
            alpha_s = np.where(xi_w > xi_w_last, alpha_s + 0.1, alpha_s - 0.1)

            # 更新帧谱
            g = np.minimum(xi_w, 1)
        else:
            # 更新噪声谱
            noise_mu = (1 - noise_mu) * noise_mu + noise_mu * np.minimum(frame_energy[i] / noise_mu, 1)
            noise_energy = np.append(noise_energy[1:], frame_energy[i])
            noise_energy_db = 10 * np.log10(noise_energy)
            xi_w_last = xi_w
            g = np.ones(frame_len)

        # 应用谱增益
        frames_fft[i] *= g

    # 合成信号
    stft = np.zeros((num_frames, N), dtype=np.complex64)
    stft[:, :N//2+1] = frames_fft
    stft[:, N//2+1:] = np.flip(np.conj(frames_fft[:, 1:N//2]), axis=1)
    sig_out = signal.istft(stft, hop_length=N//2)[1]

    # 去掉前面的静音部分
    start = np.argmax(sub_speech_flag)
    sig_out = sig_out[start * N//2:]

    # 恢复幅度
    sig_out *= np.max(np.abs(sig)) / np.max(np.abs(sig_out))

    return sig_out

最后，我们可以读取一个wav文件，调用logmmse函数进行降噪，然后保存降噪后的文件：

# 读取wav文件
rate, sig = wavfile.read('test.wav')

# 降噪
sig_out = logmmse(sig, rate)

# 保存降噪后的文件
wavfile.write('test_out.wav', rate, sig_out.astype(np.int16))

注意，logmmse函数中的参数可以根据实际情况进行调整，例如：

- noise_frames：用于估计噪声能量的帧数
- SSub：语音信号的能量与噪声能量的比值
- alpha：平滑系数
- over_sub：连续多少帧使用xi_w_last+1进行谱增益调整
- noise_mu：初始噪声能量估计值