MATLAB代码丨信号处理：对Python中Librosa库部分函数的重现

目录

Librosa：用于信号处理的Python库

对Librosa部分函数的MATLAB重现

谱重心（Centroid）

滚降频率（Rolloff）

均方根（RMS）

过零率（ZeroCrossingRate）

---

Librosa：用于信号处理的Python库

官方介绍：A python package for music and audio analysis

网址：librosa --- librosa 0.11.0 documentation

代码网址：https://github.com/librosa/librosa

安装：https://github.com/librosa/librosa/blob/main/README.md

对Librosa部分函数的MATLAB重现

常用的信号处理特征及指标有均方根（RMS）、谱重心、谱宽度、过零率、滚降频率、梅尔频率倒谱系数（MFCCs）等等。

MATLAB中也有一些用于信号处理的拓展包，例如Audio Toolbox、Signal Processing Toolbox等等，其中是否有和Librosa函数直接对标的函数暂时没有进行调研。

这里对于几个Librosa中的函数进行Matlab重现，并且对于IRMAS数据集中的部分音频进行测试，与Librosa提取得到结果进行比对。

谱重心（Centroid）

Librosa库中调用

spec_cent = librosa.feature.spectral_centroid(y=y, sr=sr)

MATLAB复现

function centroid = getCentroid(y, fs)
    framelength = 2048;
    hoplength = 512;
    nfft = 2048;
    if size(y, 2) > 1
        y = y(:, 1);
    end
    numFrames = ceil((length(y) - framelength) / hoplength) + 1;
    padLength = (numFrames - 1) * hoplength + framelength - length(y);
    y = [y; zeros(padLength, 1)];
    centroid = zeros(numFrames, 1);
    freqs = (0:(nfft/2)) * (fs / nfft);
    for i = 1:numFrames
        startIdx = (i - 1) * hoplength + 1;
        endIdx = startIdx + framelength - 1;
        frame = y(startIdx:endIdx);
        frame = frame .* hanning(framelength);
        spectrum = abs(fft(frame, nfft));
        spectrum = spectrum(1:nfft/2+1);
        numerator = sum(freqs.*spectrum);
        denominator = sum(spectrum);
        if denominator > 0
            centroid(i) = numerator / denominator;
        else
            centroid(i) = 0;
        end
    end
end

验证结果可视化

滚降频率（Rolloff）

Librosa库中调用

rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr)

MATLAB复现

function rolloff = getRolloff(y, fs, roll_percent)
    if nargin < 3
        roll_percent = 0.85;
    end
    if roll_percent < 0 || roll_percent >= 1
        error('roll_percent must lie in the range (0, 1)');
    end
    if size(y, 2) > 1
        y = mean(y, 2);
    end
    n_fft = 2048;
    hop_length = 512;
    window = hann(n_fft);
    [S, f, t] = spectrogram(y, window, n_fft-hop_length, n_fft, fs);
    S = abs(S);
    total_energy = cumsum(S, 1);
    if isempty(f)
        f = linspace(0, fs/2, n_fft/2 + 1);
    end
    threshold = roll_percent * total_energy(end, :);
    rolloff = NaN(1, length(t));
    for i = 1:length(t)
        [~, ind] = find(total_energy(:, i) >= threshold(:, i), 1, 'first');
        if ~isempty(ind)
            rolloff(i) = f(ind);
        end
    end
end

验证结果可视化

均方根（RMS）

Librosa库中调用

rms = librosa.feature.rms(y=y)

MATLAB复现

function rms_vals = getRMS(y, fs)
    frame_length = 2048;
    hop_length = 512;
    if size(y, 2) > 1
        y = mean(y, 2);
    end
    y = y / max(abs(y));
    num_frames = ceil((length(y) - frame_length) / hop_length) + 1;
    padded_length = (num_frames - 1) * hop_length + frame_length;
    y_padded = [y; zeros(padded_length - length(y), 1)];
    rms_vals = zeros(1, num_frames);
    for i = 1:num_frames
        frame_start = (i - 1) * hop_length + 1;
        frame_end = frame_start + frame_length - 1;
        frame = y_padded(frame_start:frame_end);
        rms_vals(i) = sqrt(mean(frame.^2));
    end
end

验证结果可视化

过零率（ZeroCrossingRate）

Librosa库中调用

zero_cross = librosa.feature.zero_crossing_rate(y=y)

MATLAB复现

function zcr = getZeroCrossing_rate(y)
    frame_length = 2048;
    hop_length = 512;
    center = true;
    if size(y, 2) > 1
        y = mean(y, 2);
    end
    if ~isvector(y)
        error('y should be a vector (signal)');
    end
    if center
        padding_length = floor(frame_length / 2);
        y = [repmat(y(1), padding_length, 1); y; repmat(y(end), padding_length, 1)];
    end
    frames = buffer(y, frame_length, frame_length - hop_length);
    num_frames = size(frames, 2);
    zcr = zeros(1, num_frames);
    for i = 1:num_frames
        frame = frames(:, i);
        crossings = sum(abs(diff(sign(frame))) > 0);
        zcr(i) = crossings / frame_length;
    end
end

验证结果可视化

---

END