Massively Multilingual Speech
lastmod 2023-05-24created 2023-05-242 edits
A speech technology project for 1000+ languages released by Meta
Features
- A project to build a single speech model supporting thousands of languages, going beyond existing speech recognition models that support only about 100 languages, toward the 7000+ languages spoken on Earth
- Provides multilingual speech recognition (ASR) and speech synthesis (TTS) models for 1100+ languages, language identification (LID) models for 4000+ languages, pretrained models for 1400+ languages, and more
- Aims to let people access and use information on their devices in the language they want
- As part of the project, built a dataset of New Testament readings averaging 32 hours per language for 1100 languages, and expanded the supported languages to 4000 using various unlabeled recordings of Christian texts
Downloadable model files
- Pretraied models: MMS-300M (3.5GB) and MMS-1B (10GB)
- ASR speech recognition models and dictionaries released: MMS-1B:FL102 (102 languages, 4.5GB), MMS-1B:L1107 (1107 languages, 13GB), MMS-1B-all (1162 languages, 13.7GB)
- TTS speech synthesis models: generator and vocabulary files for each of 1107 languages
- LID language identification models: models and dictionaries for 126, 256, 512, 1024, 2048, and 4017 languages