The authors used a self-teacher speech language model to generate translations of complete utterances and fixed-duration prefixes of the same waveform.
Adapting a speech language model with self-generated prefix supervision improves simultaneous translation quality-latency trade-offs.
Training on target prefixes derived from partial-waveform translations enables flicker-free streaming without transcripts, with multi-turn append-only decoding outperforming single-turn forced-prefix protocols.
Big Tech
Hieu Hoang · Amittai Axelrod
Microsoft
Research Digest··3 min read
Thread:Latent Speech Reasoning
The authors adapted a full-utterance speech language model to simultaneous translation by creating target prefixes from its own translations of partial waveforms.
Why this paper
From Microsoft · Part of Latent Speech Reasoning, now 3 papers
In one line
Prefix learning from partial speech improves quality-latency trade-offs in simultaneous translation, especially with multi-turn decoding and confidence thresholds.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§