omniASR CTC 1B Fine-tuned on Kazakh
Fine-tuned version of omniASR CTC 1B on Kazakh speech data.
Training Data
- KSC2 (Kazakh Speech Corpus 2)
- FLEURS kk_kz
Training Details
- Framework: fairseq2 / omnilingual-asr
- Steps: 20,000
- Learning rate: 5e-5
- Augmentation: white noise (40% probability, σ=0.005) + gain jitter (50%, 0.95-1.05)
Results
| Model | WER (dev) |
|---|---|
| Base omniASR 1B | 51.4% |
| Fine-tuned (this model) | 23.92% |
Usage
Load with fairseq2 / omnilingual-asr framework.
Model tree for Zhxshy/omni-ctc-1b-kazakh
Base model
facebook/omniASR-CTC-1B