Output mel spectrograms for python and CPP don't match

[sahi1422]

Python Code

samples = np.full(16_000 * 5, 0.2)
mel_spec = librosa.feature.melspectrogram(y=samples, sr=16_000, n_fft=1024, hop_length=512, n_mels=128)

print('shape: ', mel_spec.shape)
print('sum: ', mel_spec.sum())
print('min: ', mel_spec.min())
print('max: ', mel_spec.max())

Python Output

shape:  (128, 157)
sum:  11806.58392107923
min:  7.865755998623591e-34
max:  89.1440147017634

---

CPP Code

    int sr = 16000;
    int n_fft = 1024;
    int n_hop = 512;
    int n_mel = 128;
    int fmin = 0.0;
    int fmax = sr / 2.0;

    std::vector<float> x(sr * 5, 0.2);

    std::vector<std::vector<float>> mels = librosa::Feature::melspectrogram(x, sr, n_fft, n_hop, "hann", true,
                                                                            "reflect",
                                                                            2.f, n_mel, fmin, fmax);

    std::cout << "shape: [" << mels.size() << "," << mels[0].size() << "]" << std::endl;

    double sums = 0;
    float maxi = -INT_MAX, mini = INT_MAX;
    for(auto &arr: mels) {
        for(auto k: arr) {
            sums += k;
            maxi = max(maxi, k);
            mini = min(mini, k);
        }
    }
    cout << "sums: " << sums << "\n";
    cout << "mini: " << mini << "\n";
    cout << "maxi: " << maxi << "\n";

CPP output

shape: [157,128]
sums: 11761.6
mini: 5.23758e-16
maxi: 74.9149

[shreks7]

I have the same issue, the outputs don't match.

[abdr17]

Can you tell me some resources to study basics of audio signal processing like concepts of STFT, MelSpectrogram, etc(code oriented).

[xiehengye]

Python Code

samples = np.full(16_000 * 5, 0.2)
mel_spec = librosa.feature.melspectrogram(y=samples, sr=16_000, n_fft=1024, hop_length=512, n_mels=128)

print('shape: ', mel_spec.shape)
print('sum: ', mel_spec.sum())
print('min: ', mel_spec.min())
print('max: ', mel_spec.max())

Python Output

shape:  (128, 157)
sum:  11806.58392107923
min:  7.865755998623591e-34
max:  89.1440147017634

CPP Code

    int sr = 16000;
    int n_fft = 1024;
    int n_hop = 512;
    int n_mel = 128;
    int fmin = 0.0;
    int fmax = sr / 2.0;

    std::vector<float> x(sr * 5, 0.2);

    std::vector<std::vector<float>> mels = librosa::Feature::melspectrogram(x, sr, n_fft, n_hop, "hann", true,
                                                                            "reflect",
                                                                            2.f, n_mel, fmin, fmax);

    std::cout << "shape: [" << mels.size() << "," << mels[0].size() << "]" << std::endl;

    double sums = 0;
    float maxi = -INT_MAX, mini = INT_MAX;
    for(auto &arr: mels) {
        for(auto k: arr) {
            sums += k;
            maxi = max(maxi, k);
            mini = min(mini, k);
        }
    }
    cout << "sums: " << sums << "\n";
    cout << "mini: " << mini << "\n";
    cout << "maxi: " << maxi << "\n";

CPP output

shape: [157,128]
sums: 11761.6
mini: 5.23758e-16
maxi: 74.9149

I found the same problem, did you solve it?

[ewan-xu]

setting pad_mode = "reflect" in python code like this:

samples = np.full(16000 * 5, 0.2)
mel_spec = librosa.feature.melspectrogram(y=samples, sr=16000, n_fft=1024, hop_length=512, n_mels=128,pad_mode='reflect')

print(mel_spec.T)
print('shape: ', mel_spec.shape)
print('sum: ', mel_spec.sum())
print('min: ', mel_spec.min())
print('max: ', mel_spec.max())

python output:

shape:  (128, 157)
sum:  11761.64503417969
min:  1.1692227111050498e-33
max:  74.91493652343752