Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,60 @@
+---
+language:
+- nya
+license: mit
+tags:
+- asr
+- speech
+- chichewa
+- nyanja
+- wav2vec2
+- ctc
+- fairseq2
+datasets:
+- CLEAR-Global/Chichewa-Synthetic-ASR-Dataset
+metrics:
+- wer
+---
+
+# omniASR CTC 300M_v2 — fine-tuned on Chichewa/Nyanja (synthetic only)
+
+Fine-tuned checkpoint of [OmniLingual ASR CTC 300M_v2](https://github.com/facebookresearch/omnilingual-asr)
+on the CLEAR Global Chichewa Synthetic ASR Dataset (374,576 examples, ~550 hours of TTS-synthesised speech).
+Training used the fairseq2 wav2vec2 ASR recipe for 5,000 steps with the encoder frozen for the first
+1,000 steps. The FLEURS Chichewa dev split was used for validation (the synthetic dataset has no dev split).
+Tokenizer: `omniASR_tokenizer_written_v2`.
+
+## Results
+
+| Dataset | WER (%) | CER (%) |
+|---|---:|---:|
+| FLEURS Chichewa test | 58.70 | 14.92 |
+| Zambezi Voice Nyanja test | 70.47 | 16.56 |
+
+For reference, the zero-shot omniASR CTC 300M_v2 baseline reaches 51.92% WER on FLEURS and 71.87%
+on Zambezi Voice. Synthetic-only fine-tuning is worse than zero-shot on FLEURS due to acoustic
+domain mismatch between TTS and natural speech, while providing a small improvement on the
+out-of-domain Zambezi Voice benchmark. See also the FLEURS fine-tuned variant
+(`ai4good-labyrinth/omniASR-CTC-300M-v2-chichewa`) which achieves 36.53% on FLEURS.
+
+## Training details
+
+- **Base model:** omniASR_CTC_300M_v2
+- **Tokenizer:** omniASR_tokenizer_written_v2
+- **Training data:** CLEAR-Global/Chichewa-Synthetic-ASR-Dataset train split (374,576 examples, ~550 h)
+- **Steps:** 5,000 (best checkpoint at step 2,500 by FLEURS dev WER)
+- **Dev WER at best checkpoint:** 59.12% (FLEURS dev)
+- **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`.