Add zerobench_eval: official standalone scorer (pre-generated wavs in, metrics out)

#2
ZeroWeight AI org

Standalone scorer that never loads a TTS model.

python -m zerobench_eval manifest --out manifest.jsonl
python -m zerobench_eval score --wav_dir my_wavs/ --name MyModel

  • two ASRs (whisper-large-v3 + PhoWhisper-large), min taken
  • per-span reference expansion; three policies reported (strict/norm/robust)
  • tone-only variants never admitted (Vietnamese tone is phonemic)
  • UTMOS seeded for reproducibility
  • test_references.py pins both directions of the WER policy
zeroweightai changed pull request status to closed

Sign up or log in to comment