File size: 338 Bytes
0850906 |
1 2 3 4 5 6 |
# Stage-1 RVQ for Twi (Whisper encoder features)
- Single RVQ wrapper (no LayerNorms) with `ema_counts` buffer for parity
- Files: `config_stage1.json`, `rvq_final.pt`, `rvq_best.pt` (if any), `rvq_averaged_full.pt`, `rvq_wrapper.py`
- Decoding uses Whisper generate with timestamps disabled and no forced decoder ids (transcribe mode).
|