MMS
A large-scale multilingual speech model released by Meta that employs a pure encoder architecture. It supports speech recognition and synthesis across more than 1,000 languages.
Timeline (last 90 days)
Multilingual self-supervised models such as XLSR and MMS extend the self-supervised approach to dozens or even thousands of languages.
The study employs three ASR architectures: MMS (encoder-only), Whisper (encoder-decoder), and Qwen3-ASR (LLM-based).
MMS is an encoder-only architecture.
If ALIA can focus on high-value vertical scenarios (such as healthcare, agriculture, and finance) to build specialized datasets, it will be complementary to rather than competitive with broad-coverage models like MMS
Meta's MMS model is based on the wav2vec 2.0 architecture and achieves speech recognition coverage by using religious texts (primarily multilingual translations of the Bible) as weakly supervised data sources
Meta's MMS (Massively Multilingual Speech) project has already covered pre-trained models for more than 1,100 languages.
Multilingual speech recognition models such as OpenAI Whisper, Google's USM, and Meta's MMS have lowered the technical barrier for speech-to-text in multilingual mixed-language scenarios.
OpenAI Whisper was trained on 99 languages, while Meta's MMS covers more than 1,100 languages.
All Facts (8)
Meta's MMS (Massively Multilingual Speech) project has already covered pre-trained models for more than 1,100 languages.
60%UnverifiedOpenAI Whisper was trained on 99 languages, while Meta's MMS covers more than 1,100 languages.
60%UnverifiedMultilingual self-supervised models such as XLSR and MMS extend the self-supervised approach to dozens or even thousands of languages.
50%UnverifiedThe study employs three ASR architectures: MMS (encoder-only), Whisper (encoder-decoder), and Qwen3-ASR (LLM-based).
50%UnverifiedMMS is an encoder-only architecture.
50%UnverifiedMeta's MMS model is based on the wav2vec 2.0 architecture and achieves speech recognition coverage by using religious texts (primarily multilingual translations of the Bible) as weakly supervised data sources
50%UnverifiedIf ALIA can focus on high-value vertical scenarios (such as healthcare, agriculture, and finance) to build specialized datasets, it will be complementary to rather than competitive with broad-coverage models like MMS
50%UnverifiedMultilingual speech recognition models such as OpenAI Whisper, Google's USM, and Meta's MMS have lowered the technical barrier for speech-to-text in multilingual mixed-language scenarios.
50%