Skip to content

O1-Sound research checkpoint — 10-language MSWC run (English usable, multilingual open)

Latest

Choose a tag to compare

@everest-an everest-an released this 20 Aug 09:27
· 1 commit to main since this release

O1-Sound research checkpoint (2026-08-18)

Honest status — read first:

  • English is the only language with meaningful positive data (41 clips): FRR 0.098 at FAR 0.133 in the 10-language run (earlier English-only run: FRR 0.146 @ FAR 0.046)
  • Multilingual is not viable yet: most languages have 2-13 positive clips vs 400 negatives; per-language FRR ranges 0.25-1.0 — bounded by the worst language, not the mean
  • Research checkpoint, not a production wake-word model. Full record in RESULTS.md.

Contents: o1sound.pt (4.8 MB) · o1sound.onnx (fp32 5.04 MB streaming export) · o1sound.metrics.json (per-language dev metrics)

License: MIT