多窗谱谱减法 (Multi-taper Spectral Subtraction) 语音去噪实现，这是一种比传统谱减法更鲁棒的语音增强方法。
MATLAB 实现
2 主程序：多窗谱谱减法语音去噪
matlab
%% 多窗谱谱减法语音去噪
% 方法：Multi-taper Spectral Subtraction
% 特点：减少音乐噪声，提高语音质量
% 适用：非平稳噪声环境下的语音增强
clear; clc; close all;
warning off;
%% 参数设置
frame_len = 256; % 帧长 (16kHz采样率下约16ms)
frame_shift = 128; % 帧移 (50%重叠)
fs = 16000; % 采样率
n_tapers = 4; % 多窗谱使用的Slepian窗数量
alpha = 2.5; % 过减因子 (通常2-4)
beta = 0.02; % 谱下限参数 (防止音乐噪声)
gamma = 0.8; % 平滑因子
noise_frames = 10; % 用于噪声估计的初始静音帧数
%% 读取语音文件
[noisy_speech, fs] = audioread('noisy_speech.wav');
if size(noisy_speech, 2) > 1
noisy_speech = mean(noisy_speech, 2); % 转为单声道
end
% 如果有纯净语音用于评估
try
[clean_speech, ~] = audioread('clean_speech.wav');
has_clean = true;
catch
has_clean = false;
end
fprintf('=== 多窗谱谱减法语音去噪 ===\n');
fprintf('采样率: %d Hz\n', fs);
fprintf('帧长: %d 样本 (%.1f ms)\n', frame_len, frame_len/fs*1000);
fprintf('多窗谱窗数: %d\n', n_tapers);
fprintf('过减因子 α: %.1f\n', alpha);
fprintf('谱下限 β: %.3f\n', beta);
%% 多窗谱谱减法处理
enhanced_speech = multi_taper_spectral_subtraction(...
noisy_speech, fs, frame_len, frame_shift, ...
n_tapers, alpha, beta, gamma, noise_frames);
%% 评估指标计算
if has_clean
% 计算SNR改善
snr_before = 10*log10(sum(clean_speech.^2)/sum((noisy_speech-clean_speech).^2));
snr_after = 10*log10(sum(clean_speech.^2)/sum((enhanced_speech-clean_speech).^2));
fprintf('\n=== 客观评估 ===\n');
fprintf('输入SNR: %.2f dB\n', snr_before);
fprintf('输出SNR: %.2f dB\n', snr_after);
fprintf('SNR改善: %.2f dB\n', snr_after - snr_before);
% 计算PESQ分数（需要Voicebox工具箱）
try
pesq_score = pesq(clean_speech, enhanced_speech, fs);
fprintf('PESQ分数: %.3f\n', pesq_score);
catch
fprintf('PESQ分数: 未计算（需要Voicebox工具箱）\n');
end
end
%% 保存结果
audiowrite('enhanced_speech.wav', enhanced_speech, fs);
fprintf('增强语音已保存为: enhanced_speech.wav\n');
%% 可视化结果
visualize_results(noisy_speech, enhanced_speech, clean_speech, fs, has_clean);
%% ========== 核心函数 ==========
function enhanced = multi_taper_spectral_subtraction(...
noisy, fs, frame_len, frame_shift, n_tapers, alpha, beta, gamma, noise_frames)
% 多窗谱谱减法主函数
% 分帧
frames = enframe(noisy, frame_len, frame_shift);
n_frames = size(frames, 1);
% 生成Slepian窗（多窗谱的核心）
slepian_windows = dpss(frame_len, n_tapers);
% 初始化
enhanced_frames = zeros(size(frames));
noise_psd = zeros(frame_len, 1);
prev_gain = ones(frame_len, 1);
% 噪声估计（使用前几帧）
for i = 1:noise_frames
frame = frames(i, :)';
noise_psd = noise_psd + abs(fft(frame)).^2;
end
noise_psd = noise_psd / noise_frames;
% 处理每一帧
for i = 1:n_frames
frame = frames(i, :)';
% 多窗谱估计
multi_taper_spec = zeros(frame_len, 1);
for taper = 1:n_tapers
windowed_frame = frame .* slepian_windows(taper, :)';
spec = fft(windowed_frame);
multi_taper_spec = multi_taper_spec + abs(spec).^2;
end
multi_taper_spec = multi_taper_spec / n_tapers;
% 计算增益函数（谱减法）
gain = max(1 - alpha * sqrt(noise_psd ./ (multi_taper_spec + eps)), beta);
% 时间平滑
gain = gamma * prev_gain + (1 - gamma) * gain;
prev_gain = gain;
% 应用增益
enhanced_spec = gain .* fft(frame);
% IFFT得到时域信号
enhanced_frame = real(ifft(enhanced_spec));
enhanced_frames(i, :) = enhanced_frame';
% 更新噪声估计（语音活动检测）
if i > noise_frames
noise_psd = 0.95 * noise_psd + 0.05 * multi_taper_spec;
end
end
% 重叠相加重构语音
enhanced = overlap_add(enhanced_frames, frame_shift);
end
function frames = enframe(x, win_len, shift_len)
% 分帧函数
n_frames = floor((length(x) - win_len) / shift_len) + 1;
frames = zeros(n_frames, win_len);
for i = 1:n_frames
start_idx = (i-1)*shift_len + 1;
end_idx = start_idx + win_len - 1;
frames(i, :) = x(start_idx:end_idx);
end
end
function x = overlap_add(frames, shift_len)
% 重叠相加重构
[n_frames, frame_len] = size(frames);
x_len = (n_frames-1)*shift_len + frame_len;
x = zeros(x_len, 1);
for i = 1:n_frames
start_idx = (i-1)*shift_len + 1;
end_idx = start_idx + frame_len - 1;
x(start_idx:end_idx) = x(start_idx:end_idx) + frames(i, :)';
end
end
function windows = dpss(N, K)
% 生成离散扁长球序列（DPSS）窗
% N: 窗长度
% K: 窗数量
% 返回: K×N 的窗矩阵
% 时间带宽积
NW = 3;
% 计算DPSS
[windows, ~] = dpss(N, NW, K);
% 归一化
for i = 1:K
windows(i, :) = windows(i, :) / norm(windows(i, :));
end
end
%% ========== 可视化函数 ==========
function visualize_results(noisy, enhanced, clean, fs, has_clean)
% 可视化结果
% 时间轴
t = (0:length(noisy)-1)/fs;
figure('Position', [100, 100, 1200, 800]);
% 波形图
subplot(3, 2, 1);
plot(t, noisy, 'b', 'LineWidth', 1);
xlabel('时间 (s)'); ylabel('幅度');
title('带噪语音波形');
grid on; axis tight;
subplot(3, 2, 2);
plot(t, enhanced, 'r', 'LineWidth', 1);
xlabel('时间 (s)'); ylabel('幅度');
title('增强语音波形');
grid on; axis tight;
if has_clean
subplot(3, 2, 3);
plot(t, clean, 'g', 'LineWidth', 1);
xlabel('时间 (s)'); ylabel('幅度');
title('纯净语音波形');
grid on; axis tight;
end
% 语谱图
subplot(3, 2, 4);
spectrogram(noisy, hamming(256), 128, 512, fs, 'yaxis');
title('带噪语音语谱图');
colorbar;
subplot(3, 2, 5);
spectrogram(enhanced, hamming(256), 128, 512, fs, 'yaxis');
title('增强语音语谱图');
colorbar;
if has_clean
subplot(3, 2, 6);
spectrogram(clean, hamming(256), 128, 512, fs, 'yaxis');
title('纯净语音语谱图');
colorbar;
end
% 频谱对比
figure('Position', [100, 100, 800, 600]);
% 计算平均频谱
noisy_spec = abs(fft(enframe(noisy, 1024, 512)));
enhanced_spec = abs(fft(enframe(enhanced, 1024, 512)));
freq = (0:size(noisy_spec, 2)-1)*fs/size(noisy_spec, 2);
subplot(2, 1, 1);
plot(freq(1:512), mean(noisy_spec(:, 1:512), 1), 'b', 'LineWidth', 1.5);
hold on;
plot(freq(1:512), mean(enhanced_spec(:, 1:512), 1), 'r', 'LineWidth', 1.5);
xlabel('频率 (Hz)'); ylabel('幅度');
title('平均频谱对比');
legend('带噪语音', '增强语音');
grid on; axis tight;
% 频谱差异
subplot(2, 1, 2);
diff_spec = mean(enhanced_spec(:, 1:512), 1) - mean(noisy_spec(:, 1:512), 1);
plot(freq(1:512), diff_spec, 'g', 'LineWidth', 1.5);
xlabel('频率 (Hz)'); ylabel('幅度差');
title('频谱差异 (增强 - 带噪)');
grid on; axis tight;
sgtitle('多窗谱谱减法语音增强结果');
end2 改进的变参数多窗谱谱减法
matlab
%% 改进版：自适应多窗谱谱减法
function enhanced = adaptive_multi_taper_ss(noisy, fs)
% 自适应参数的多窗谱谱减法
% 参数
frame_len = 256;
frame_shift = 128;
n_tapers = 4;
% 分帧
frames = enframe(noisy, frame_len, frame_shift);
n_frames = size(frames, 1);
% 生成Slepian窗
slepian_windows = dpss(frame_len, n_tapers);
% 初始化
enhanced_frames = zeros(size(frames));
noise_psd = zeros(frame_len, 1);
prev_gain = ones(frame_len, 1);
% 初始噪声估计
noise_est_frames = min(10, round(n_frames/10));
for i = 1:noise_est_frames
frame = frames(i, :)';
noise_psd = noise_psd + abs(fft(frame)).^2;
end
noise_psd = noise_psd / noise_est_frames;
% 处理每一帧
for i = 1:n_frames
frame = frames(i, :)';
% 多窗谱估计
multi_taper_spec = zeros(frame_len, 1);
for taper = 1:n_tapers
windowed_frame = frame .* slepian_windows(taper, :)';
spec = fft(windowed_frame);
multi_taper_spec = multi_taper_spec + abs(spec).^2;
end
multi_taper_spec = multi_taper_spec / n_tapers;
% 计算信噪比
snr_est = 10*log10(mean(multi_taper_spec) / mean(noise_psd));
% 自适应过减因子
if snr_est > 20
alpha = 1.0; % 高信噪比，少减
elseif snr_est > 10
alpha = 2.0;
elseif snr_est > 0
alpha = 3.0;
else
alpha = 4.0; % 低信噪比，多减
end
% 自适应谱下限
beta = 0.01 * (1 - exp(-snr_est/10));
% 计算增益函数
gain = max(1 - alpha * sqrt(noise_psd ./ (multi_taper_spec + eps)), beta);
% 时间平滑
gamma = 0.7 + 0.2 * exp(-snr_est/5); % 自适应平滑因子
gain = gamma * prev_gain + (1 - gamma) * gain;
prev_gain = gain;
% 应用增益
enhanced_spec = gain .* fft(frame);
% IFFT
enhanced_frame = real(ifft(enhanced_spec));
enhanced_frames(i, :) = enhanced_frame';
% 更新噪声估计（语音活动检测）
vad_threshold = 0.1;
if mean(multi_taper_spec) < mean(noise_psd) * (1 + vad_threshold)
noise_psd = 0.99 * noise_psd + 0.01 * multi_taper_spec;
end
end
% 重构语音
enhanced = overlap_add(enhanced_frames, frame_shift);
end3 多窗谱谱减法与传统方法对比
matlab
%% 多窗谱谱减法与传统方法对比
function compare_methods(noisy, fs)
% 对比传统谱减法、维纳滤波和多窗谱谱减法
frame_len = 256;
frame_shift = 128;
% 方法1：传统谱减法
enhanced_ss = traditional_spectral_subtraction(noisy, fs, frame_len, frame_shift);
% 方法2：维纳滤波
enhanced_wiener = wiener_filtering(noisy, fs, frame_len, frame_shift);
% 方法3：多窗谱谱减法
enhanced_mtss = multi_taper_spectral_subtraction(...
noisy, fs, frame_len, frame_shift, 4, 2.5, 0.02, 0.8, 10);
% 可视化对比
figure('Position', [100, 100, 1400, 800]);
t = (0:length(noisy)-1)/fs;
% 波形对比
subplot(2, 2, 1);
plot(t, noisy, 'b', 'LineWidth', 1);
title('带噪语音');
xlabel('时间 (s)'); ylabel('幅度');
grid on;
subplot(2, 2, 2);
plot(t, enhanced_ss, 'r', 'LineWidth', 1);
title('传统谱减法');
xlabel('时间 (s)'); ylabel('幅度');
grid on;
subplot(2, 2, 3);
plot(t, enhanced_wiener, 'g', 'LineWidth', 1);
title('维纳滤波');
xlabel('时间 (s)'); ylabel('幅度');
grid on;
subplot(2, 2, 4);
plot(t, enhanced_mtss, 'm', 'LineWidth', 1);
title('多窗谱谱减法');
xlabel('时间 (s)'); ylabel('幅度');
grid on;
% 语谱图对比
figure('Position', [100, 100, 1400, 800]);
subplot(2, 2, 1);
spectrogram(noisy, hamming(256), 128, 512, fs, 'yaxis');
title('带噪语音');
subplot(2, 2, 2);
spectrogram(enhanced_ss, hamming(256), 128, 512, fs, 'yaxis');
title('传统谱减法');
subplot(2, 2, 3);
spectrogram(enhanced_wiener, hamming(256), 128, 512, fs, 'yaxis');
title('维纳滤波');
subplot(2, 2, 4);
spectrogram(enhanced_mtss, hamming(256), 128, 512, fs, 'yaxis');
title('多窗谱谱减法');
% 计算客观指标
fprintf('\n=== 方法对比 ===\n');
% 计算音乐噪声（高频能量比）
music_noise_ss = calculate_music_noise(enhanced_ss, fs);
music_noise_wiener = calculate_music_noise(enhanced_wiener, fs);
music_noise_mtss = calculate_music_noise(enhanced_mtss, fs);
fprintf('音乐噪声指标（越小越好）：\n');
fprintf('传统谱减法: %.4f\n', music_noise_ss);
fprintf('维纳滤波: %.4f\n', music_noise_wiener);
fprintf('多窗谱谱减法: %.4f\n', music_noise_mtss);
end
function music_noise = calculate_music_noise(signal, fs)
% 计算音乐噪声指标（高频能量比）
[frames, ~] = buffer(signal, 256, 128, 'nodelay');
music_noise = 0;
for i = 1:size(frames, 2)
frame = frames(:, i);
spec = abs(fft(frame));
freq = (0:length(spec)-1)*fs/length(spec);
% 高频能量（>4kHz）与总能量比
high_freq_energy = sum(spec(freq > 4000));
total_energy = sum(spec);
music_noise = music_noise + high_freq_energy / (total_energy + eps);
end
music_noise = music_noise / size(frames, 2);
end4 实际应用示例
matlab
%% 实际应用示例：处理真实录音
function practical_example()
% 录制或加载语音
fs = 16000;
duration = 5; % 录制5秒
fprintf('请说话...（录制%d秒）\n', duration);
recObj = audiorecorder(fs, 16, 1);
recordblocking(recObj, duration);
noisy = getaudiodata(recObj, 'double');
% 添加模拟噪声（白噪声）
noise_level = 0.05;
noisy = noisy + noise_level * randn(size(noisy));
% 应用多窗谱谱减法
enhanced = multi_taper_spectral_subtraction(...
noisy, fs, 256, 128, 4, 2.5, 0.02, 0.8, 10);
% 播放结果
fprintf('播放带噪语音...\n');
sound(noisy, fs);
pause(duration + 1);
fprintf('播放增强语音...\n');
sound(enhanced, fs);
% 保存结果
audiowrite('recorded_noisy.wav', noisy, fs);
audiowrite('recorded_enhanced.wav', enhanced, fs);
fprintf('录音已保存\n');
end
%% 批量处理文件夹中的音频文件
function batch_process(folder_path)
files = dir(fullfile(folder_path, '*.wav'));
for i = 1:length(files)
filename = fullfile(folder_path, files(i).name);
[noisy, fs] = audioread(filename);
% 增强处理
enhanced = multi_taper_spectral_subtraction(...
noisy, fs, 256, 128, 4, 2.5, 0.02, 0.8, 10);
% 保存增强后的文件
[path, name, ext] = fileparts(filename);
enhanced_filename = fullfile(path, [name '_enhanced' ext]);
audiowrite(enhanced_filename, enhanced, fs);
fprintf('处理完成: %s -> %s\n', files(i).name, [name '_enhanced' ext]);
end
end参考代码 多窗谱谱减法，语音去噪 www.youwenfan.com/contentcst/77736.html
参数调优指南
| 参数 | 推荐值 | 作用 | 调整建议 |
|---|---|---|---|
n_tapers |
3-5 | 多窗谱窗数量 | 增加可减少频谱估计方差，但增加计算量 |
alpha |
2.0-4.0 | 过减因子 | 噪声越大，值应越大 |
beta |
0.01-0.05 | 谱下限 | 防止音乐噪声，太小会产生残留噪声 |
gamma |
0.7-0.9 | 时间平滑因子 | 越大越平滑，但可能丢失瞬态信息 |
noise_frames |
5-20 | 噪声估计帧数 | 根据静音段长度调整 |
算法优势
- 减少音乐噪声：多窗谱估计降低频谱估计的方差
- 更好的语音保留：自适应参数调整
- 计算效率：相比传统谱减法增加有限计算量
- 鲁棒性：对非平稳噪声有更好的适应性
性能评估指标
matlab
%% 综合评估函数
function evaluate_performance(clean, noisy, enhanced, fs)
% 计算多种客观指标
% SNR改善
snr_before = 10*log10(sum(clean.^2)/sum((noisy-clean).^2));
snr_after = 10*log10(sum(clean.^2)/sum((enhanced-clean).^2));
% 分段信噪比 (SegSNR)
seg_snr_before = segmental_snr(clean, noisy, fs);
seg_snr_after = segmental_snr(clean, enhanced, fs);
% 对数谱距离 (LSD)
lsd = log_spectral_distance(clean, enhanced, fs);
% 语音质量感知评估 (PESQ)
try
pesq_score = pesq(clean, enhanced, fs);
catch
pesq_score = NaN;
end
fprintf('\n=== 性能评估 ===\n');
fprintf('输入SNR: %.2f dB\n', snr_before);
fprintf('输出SNR: %.2f dB\n', snr_after);
fprintf('SNR改善: %.2f dB\n', snr_after - snr_before);
fprintf('分段SNR改善: %.2f dB\n', seg_snr_after - seg_snr_before);
fprintf('对数谱距离: %.2f\n', lsd);
if ~isnan(pesq_score)
fprintf('PESQ分数: %.3f\n', pesq_score);
end
end
function seg_snr = segmental_snr(clean, enhanced, fs)
% 计算分段信噪比
frame_len = 256;
frame_shift = 128;
frames_clean = enframe(clean, frame_len, frame_shift);
frames_enhanced = enframe(enhanced, frame_len, frame_shift);
seg_snr = 0;
n_frames = size(frames_clean, 1);
for i = 1:n_frames
signal_power = mean(frames_clean(i, :).^2);
noise_power = mean((frames_enhanced(i, :) - frames_clean(i, :)).^2);
if noise_power > 0
seg_snr = seg_snr + 10*log10(signal_power/noise_power);
end
end
seg_snr = seg_snr / n_frames;
end
function lsd = log_spectral_distance(clean, enhanced, fs)
% 计算对数谱距离
frame_len = 256;
frame_shift = 128;
frames_clean = enframe(clean, frame_len, frame_shift);
frames_enhanced = enframe(enhanced, frame_len, frame_shift);
lsd = 0;
n_frames = size(frames_clean, 1);
for i = 1:n_frames
spec_clean = abs(fft(frames_clean(i, :)'));
spec_enhanced = abs(fft(frames_enhanced(i, :)'));
% 避免零值
spec_clean = max(spec_clean, eps);
spec_enhanced = max(spec_enhanced, eps);
% 对数谱距离
lsd = lsd + sqrt(mean((10*log10(spec_clean) - 10*log10(spec_enhanced)).^2));
end
lsd = lsd / n_frames;
end使用建议
- 对于强噪声环境 ：增加
alpha到 3.5-4.0 - 对于弱噪声环境 ：减少
alpha到 1.5-2.0 - 对于音乐噪声敏感 ：减小
beta到 0.005-0.01 - 对于语音质量要求高：使用自适应版本
- 实时应用 ：减少
n_tapers到 2-3 以降低计算复杂度
这个多窗谱谱减法实现可以有效减少传统谱减法产生的音乐噪声，同时保持较好的语音质量。