Translatotron 2
Google's landmark direct speech-to-speech translator: end-to-end translated speech with the speaker's own voice preserved — deliberately engineered so it can't be repurposed for deepfakes.
About
Translatotron 2 (2021, ICML 2022, arXiv 2107.08661) is Google's second-generation neural direct speech-to-speech translation model and the first one trainable end-to-end: a speech encoder, a phoneme decoder, a duration-based mel-spectrogram synthesizer, and a single attention module tying them together. It outperformed the original Translatotron by up to +15.5 BLEU on translation quality and approached cascade-system quality on speech naturalness, while fixing the babbling and long-pause over-generation that plagued v1. Its signature idea is voice retention without any speaker embedding or ID — the translated speech keeps the source speaker's voice, and the model is deliberately restricted from generating anyone else's voice, a choice Google framed as deepfake mitigation. With a simple concatenation augmentation it can even keep each speaker's voice straight across speaker turns. It was never released as weights or an API, so it lives on as the cited proof point behind everything in direct S2ST that followed.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Up to +15.5 BLEU over the original Translatotron across three datasets (author-reported)\n- Approaches cascade-system quality on translation quality and speech naturalness (author-reported)\n- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 5/10 (TIER verified, metrics-v1). Evidence: ICML 2022 peer-reviewed paper (arXiv 2107.08661) reporting +15.5 BLEU over v1; widely cited as the direct-S2ST landmark. Strengths: first end-to-end trainable direct S2ST, voice retention without speaker embeddings, explicit anti-spoofing design. Docked for: (1) no public weights, code, or API — nothing to verify beyond the paper; (2) all quality figures are author-reported on internal datasets; (3) research is complete with no maintained release. License: not released.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 3.0/10
- Openness · license + docs
- 2.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ICML 2022 (arXiv 2107.08661); +15.5 BLEU over Translatotron v1 (author-reported)
- · First end-to-end trainable direct S2ST; voice retention without speaker embeddings, anti-spoofing design
- · No public weights, code, or API — unverifiable beyond the paper
- · Research complete; line continues in Translatotron 3 and Seamless work