Fine-tuned checkpoint of
OmniLingual ASR CTC 300M
on the FLEURS Chichewa/Nyanja train split (
google/fleurs, config
ny_mw).
Training used the fairseq2 wav2vec2 ASR recipe for 5,000 steps with the encoder
frozen for the first 1,000 steps.
For comparison, the zero-shot omniASR CTC 300M baseline reaches 54.51% WER on
FLEURS and 72.47% on Zambezi Voice. The fine-tuned model outperforms the
zero-shot 1B variant (41.36% / 63.17%) and is within 1.2 pp of the zero-shot
3B variant (35.80% / 62.10%) on both benchmarks.
