omniASR CTC 300M_v2 — fine-tuned on Chichewa/Nyanja (FLEURS + trigram 2:1)
Fine-tuned checkpoint of OmniLingual ASR CTC 300M_v2
on a mixture of FLEURS Chichewa/Nyanja (2,653 examples, ~10.87 hours) and the Chichewa Trigrams
speech-text parallel corpus (52,133 examples, ~21.75 hours — 200% of FLEURS train hours).
Training used the fairseq2 wav2vec2 ASR recipe for 5,000 steps with the encoder frozen for the
first 1,000 steps. Tokenizer: omniASR_tokenizer_written_v2.
Results
| Dataset | WER (%) | CER (%) |
|---|---|---|
| FLEURS Chichewa test | 35.56 | 7.73 |
| Zambezi Voice Nyanja test | 61.20 | 13.50 |
This model outperforms the zero-shot omniASR CTC 3B baseline (35.80% WER on FLEURS,
62.10% on Zambezi Voice) despite having 10× fewer parameters. The 1:1 trigram ratio
(ai4good-labyrinth/omniASR-CTC-300M-v2-chichewa-mixed-1p0) achieves slightly better
results (35.27% FLEURS / 60.79% Zambezi), identifying 1:1 as the optimal mixing ratio.
Training details
- Base model: omniASR_CTC_300M_v2
- Tokenizer: omniASR_tokenizer_written_v2
- Training data: FLEURS Chichewa train (2,653 examples, 10.87 h) + Chichewa Trigrams (52,133 examples, 21.75 h)
- Steps: 5,000 (best checkpoint at step 5,000 — WER still decreasing at end of training)
- Dev WER at best checkpoint: 37.05% (FLEURS dev, 305 examples)
- Learning rate: 1e-5
- Encoder frozen: first 1,000 steps
- Gradient accumulation: 4 batches
- Mixed precision: bfloat16
- Framework: fairseq2
Checkpoint format
The checkpoint is stored in fairseq2's sharded format.
checkpoint/model/pp_00/tp_00/sdp_00.pt contains the full model state
(training used a single GPU, so there is only one shard).
To load the model, use the OmniLingual ASR inference pipeline with the
omniASR_CTC_300M_v2 model card and point the checkpoint path at sdp_00.pt.