Translatotron 3

Translatotron 3 (arXiv 2305.17547, ICASSP 2024) is Google's answer to the hardest bottleneck in speech translation: paired speech data barely exists for most language pairs. It trains a direct speech-to-speech model in a fully unsupervised way from monolingual speech-text datasets alone, combining a masked autoencoder, unsupervised embedding mapping, and back-translation. On Spanish↔English it beat a cascade baseline by 18.14 BLEU on the synthesized Unpaired-Conversational set and approached supervised models on the CVSS dataset — while carrying over para-linguistic detail (pauses, speaking rates, speaker identity) that supervised pipelines have to model explicitly. Google published audio samples showing speaker-preserving translations produced without any parallel data. Like its predecessors it was never released as weights or an API, so it remains a research result rather than something you can run — but it redrew what people thought was possible without parallel corpora.

• +18.14 BLEU over an unsupervised cascade baseline on Unpaired-Conversational (author-reported)\n- Approaches supervised-model quality on English→Spanish CVSS (author-reported)\n- No independent benchmark replication found