The built-in learned model is 48 kHz-native; at 44.1 kHz it runs with a console warning and mildly detuned band edges (the preset rescales the chain's time constants, but the network's ERB band layout and training statistics assume a 24 kHz Nyquist).
Work: train a 44.1 kHz / block-256 model with the v2 recipe (tools/ml/make_dataset.py --rate 44100 needs the rate whitelisted and a spectral resample path for the 16 kHz corpus at a non-integer ratio — the current resampler already handles arbitrary ratios), evaluate with the standard meter, ship as a .munn for @model.
Depends on nothing; pairs naturally with the 16 kHz companion-model ticket.
The built-in learned model is 48 kHz-native; at 44.1 kHz it runs with a console warning and mildly detuned band edges (the preset rescales the chain's time constants, but the network's ERB band layout and training statistics assume a 24 kHz Nyquist).
Work: train a 44.1 kHz / block-256 model with the v2 recipe (
tools/ml/make_dataset.py --rate 44100needs the rate whitelisted and a spectral resample path for the 16 kHz corpus at a non-integer ratio — the current resampler already handles arbitrary ratios), evaluate with the standard meter, ship as a.munnfor@model.Depends on nothing; pairs naturally with the 16 kHz companion-model ticket.