Meta’s MMS expands speech coverage, with separate deployment checks

Meta released Massively Multilingual Speech on 22 May. Its launch files list recognition and synthesis models for 1,107 languages, plus a recognition variant covering 1,162. Language identification reaches 4,017. These are different task-specific models.

The launch README puts MMS code and weights under CC BY-NC 4.0. Its model card names researchers as intended users and calls for further risk evaluation for other fine-tuning. Language labels weren’t manually verified for every language in the MMS datasets.

The paper’s synthesis evaluation separates breadth from generalisation. Testing across all 1,107 languages uses held-out data from the same religious-reading collection and speech recognisers trained on that collection. Out-of-domain synthesis testing covers 61 languages.

About 85% of synthesis models meet the paper’s five-percent character-error threshold, measured through automatic transcription. That is a content-preservation proxy, not a human quality rating for every language.

For a voice product, language support, quality on the intended audio and permitted use need separate checks. Downloadable coverage doesn’t settle deployment suitability.

Photo: LPS.1, “Meta Platforms Headquarters Menlo Park California” (2022). CC0 1.0. Resized; cropped; responsive crops.