SpectralCentroid transform

🚀 Feature

torchaudio.transforms.SpectralCentroid

Motivation

This is a common audio transform, included in Librosa.

Pitch

A differentiable SpectralCentroid would be useful for audio neural networks.

Alternatives

Here is the librosa implementation.

Additional context

Here is a simplified implementation I have made, based upon the librosa implementation:

def spectral_centroid(y, sr, n_fft, hop_length=512):
    S = torchaudio.transforms.Spectrogram(n_fft=n_fft, hop_length=hop_length, power=1.0)(y)

#    freq = fft_frequencies(sr=sr, n_fft=n_fft)
    freq = torch.linspace(0,
                       float(sr) / 2,
                       int(1 + n_fft//2))
                       #endpoint=True)

    if freq.ndim == 1:
        freq = freq.reshape((-1, 1))

    def tl1norm(S):
        return S / torch.sum(torch.abs(S), axis=0)
        
    # Column-normalize S
    return torch.sum(freq * tl1norm(S),
                  axis=0, keepdims=True)

[jacobjwebber]

Excellent, thanks @mthrok

I will make a new merge request with optional frequency limiting, probably this weekend.

[NicolasHug]

I’ll give this a shot!