Some questions about the sample rate of speech samples about nisqa

[gabrielmittag]

Hi @liushenme, the speech samples do not need to be resampled since, as @yellowyi9527 pointed out, the length of the FFT window is adjusted according to the sampling frequency and then zero-padded to a fixed length of 4096 samples. This is done in line 2301: win_length = int(sr * win_length), where the initial win_length is 0.02 (20 ms). Because of this approach, resampling of the signals is not needed.

You are right that many of the Mel bands will be empty for a wideband or narrowband signal. However, the NISQA model predicts quality in a fullband context, so it will predict a lower score when a speech signal is missing high frequencies. For a wideband signal (up to 8 kHz), the score will on average only be slightly lower than a fullband signal. The score of a clean narrowband signal (up to 4 kHz) should on average be around a MOS of 3.8. The figure below shows the average predicted MOS vs the cutoff frequency of a lowpass.

If Librosa prints out a warning message about the empty Mel bands you can just ignore it. Basically, the model needs to know that these frequencies are empty since it expects a fullband signal.

[yellowyi9527]

He's using Libosa to read the data, and it's going to automatically go to the specified frequency。in NISQA_lib.py 2283line （y, sr = lb.load(file_path, sr=sr)）@liushenme

[liushenme]

Thanks for your reply. I understand that the author wants to resample all the speech samples to the same frequency, but after resample operation, speech samples may add many frequency dimensions with a value of 0. Is it reasonable?

And for NISQA Corpus, the speech samples are 48k sampling and their true frequency distribution is only 0-8k ( some is 0-4k ). They already have many frequency dimensions with a value of 0. I don‘t think it is reasonable to process these speech sample directly. @yellowyi9527