Preprint
SeamlessM4T: Massively Multilingual and Multimodal Machine Translation (full text)
Seamless Communication, Meta AI (arXiv), 2023 · DOI 10.48550/arXiv.2308.11596
Read the originalRead on October 17, 2026
- What it is
- The full text of the paper introducing SeamlessM4T. Its language table gives each language a resource level for speech translation into English, and the paper discusses why cascaded systems lose meaning between recognition and translation.
- How much weight it carries
- A preprint by the model's builders, so the benchmark results describe their own system. The resource levels follow the amount of supervised data for speech translation into English, not how well a language is recorded or spoken.
Facts taken from this source
- The SeamlessM4T paper lists error propagation from speech recognition into the translation model as a known weakness of cascaded systems.
“an increased likelihood of error propagation from the ASR model to the T2TT model and other subsequent models in the cascade” (Discussion of cascaded systems in the introduction)
- SeamlessM4T's language table marks Xhosa as zero-shot for speech translation into English.
“xho Xhosa Atlantic-Congo Benue-Congo Latn zero-shot Sp” (Table 5, SeamlessM4T languages)
- The SeamlessM4T paper reports a language as zero-shot for speech translation into English when it lacks the primary training data counted in its table.
“Absent such data, we report the language as zero-shot (when evaluating S2TT into English).” (Notes to the language table (resource level))
- SeamlessM4T's language table marks Zulu as a low-resource language for speech translation into English.
“zul Zulu Atlantic-Congo Benue-Congo Latn low Sp, Tx Tx” (Table 5, SeamlessM4T languages)