Repository navigation

🚨 Slim down `[audio]` to torchaudio + soundfile + soxr (drop librosa and phonemizer) - #50610

Open

eustlb wants to merge 11 commits into

mainfrom

drop-librosa-phonemizer

Open

🚨 Slim down `[audio]` to torchaudio + soundfile + soxr (drop librosa and phonemizer)#50610
eustlb wants to merge 11 commits into
mainfrom
drop-librosa-phonemizer

Commits on Nov 11, 2026
- Remove librosa from the core audio loading path
  eustlb
  authored
  c852602
- Compute parakeet / cohere_asr / nemotron_asr_streaming mel filters without librosa
  eustlb
  authored
  575969b
- Slim down the `[audio]` extra to torchaudio + soundfile + soxr
  eustlb
  authored
  92c0da5
- Build the parakeet-family mel filters with `mel_filter_bank(dtype=np.float32)`
  eustlb
  authored
  60e4fea
- Drop the standalone `[audio]` smoke test job
  eustlb
  authored
  22d9b90
- Drop the librosa comparison tests for the slaney mel filters
  eustlb
  authored
  4695350
- Remove redundant mel filter comments
  eustlb
  authored
  ccf3bc4
- Address self-review: NemotronH_Omni uses `auto`, Nyquist in `mel_filter_bank`, soxr only to resample
  eustlb
  authored
  9ab1cfe
- address review
  eustlb
  authored
  48a6d83
- keep bytes input for another PR
  eustlb
  authored
  295ac0d
- add require_audio
  eustlb
  authored
  7297598