Unable to train on 16k data

[JBloodless]

Hi,

I'm trying to retrain the model on 16khz data. I've set parameters in config.inias

sr = 16000
fft_size = 480
hop_size = 240
nb_erb = 16
nb_df = 48
min_nb_erb_freqs = 1

but at the start of the training I'm getting a bunch of

thread 'DataLoader Worker 7' panicked at 'assertion failed: `(left == right)` 
left: `16000`,
right: `48000`: bw_filterbank() assumes 48 kHz sampling rate.', libDF/src/transforms.rs:520:5

I used prepare_data.py --sr 16000 on all of my datasets tho

In transforms.rs I see

/// Estimate bandwidth by finding the highest frequency bin containing a sufficient amount of energy.
/// A minimum sampling rate of 16 kHz is assumed.

but assertion seems to be hardcoded:

pub(crate) fn estimate_bandwidth(
    input: ArrayView3<Complex32>,
    sr: usize,
    mut db_cut_off: f32,
    mut window_size: usize,
) -> usize {
    assert_eq!(sr, 48000, "bw_filterbank() assumes 48 kHz sampling rate.");

What am I doing wrong or what should I change? Thanks in advance.

[Rikorose]

The bandwidth extension only supports upsampling from down to 16khz to 48kHz. This does not make sense if your model sampling rate is 16k. Try without bandwidth extension.

[JBloodless]

Oh, thanks! Completely forgot about this setting in config.

[zzgnb]

Have you ever trained a 16k version of DF successfully? I got a significant performance gap between 16k model and the original model. Could you please give me some suggestion about configuration or some changes in code?