Overview
- Meta published a family of models from a lightweight 300M version to a 7B model, with model releases under an Apache 2.0 license.
- The system pairs a scaled wav2vec 2.0 speech encoder with a transformer decoder to enable few-shot, in-context learning for unsupported languages.
- Meta introduced an Omnilingual ASR Corpus built with partners including Mozilla’s Common Voice, Lanfrica and NaijaVoices, making resources accessible to developers and researchers.
- Company-reported results show more than 95% of high- and medium-resource languages under a 10% character error rate, but only 36% of low-resource languages meet that mark.
- Coverage highlights broad inclusion potential while noting accuracy gaps for sparse-data languages and the need for real-world validation by the community.