Preprint
Scaling Speech Technology to 1,000+ Languages
Vineel Pratap et al., arXiv, 2023
Read the originalRead on October 17, 2026
- What it is
- The full paper behind Meta's Massively Multilingual Speech (MMS) project. It compares MMS with Whisper on the FLEURS benchmark and describes the 1,107-language recognition model and its training data.
- How much weight it carries
- Peer-reviewable preprint by the developers, with a direct comparison against Whisper on a shared subset of languages. The comparison uses n-gram language models for MMS, so it is not identical to Whisper's own decoding; the authors state this.
Facts taken from this source
- Meta's MMS model reduces Whisper's word error rate by a relative 58 percent on the 54 languages that both models support in FLEURS, while supporting over 11 times as many languages.
“MMS reduces the word error rate of Whisper by a relative 58% while supporting over 11 times the number of languages.” (Section 4)
- The MMS paper says Whisper's model supports the transcription of 99 different languages.
“their model supports the transcription of 99 different languages” (Section 2)