mycroftai / sonopy Goto Github PK

After playing around with the mfcc_spec function I've noticed something and wanted some clarification on the same.

Say I have some audio file of sampling rate sr and length N * sr.
I have chosen the following parameters (matches librosa's defaults):

• window length: 2048
• window stride: 512
• n_fft: 2048
• n_filters: 128
• n_coeffs: 20

However,

1. n_fft has to be 2 * window_length - 1 for dimensionality to be preserved
2. n_filters does not work unless n_filters = window_length
3. Memory error with these default params.

When I resolve these issues by setting the params to:

• window length: 1024
• window stride: 512
• n_fft: 2047
• n_filters: 1024
• n_coeffs: 20

It results in an mfcc output of (374, 20, 1024). From what I understand, mfcc.shape[1] = n_coeffs and mfcc.shape[-1] = window_length (?) but what does mfcc.shape[0] correspond to? Furthermore, is there any way possible to try to replicate the librosa default params without running into memory issues?

Excuse me, sir, do you have sonopy mfcc based on C language?