Hibiki-Zero: process-reward simultaneous speech translation

What changed

Initialized a multistream speech translator on sentence-aligned data, then adapted GRPO with intermediate translation-quality rewards to reduce latency without word-level alignment heuristics.

Scientific scope

Initialized a multistream speech translator on sentence-aligned data, then adapted GRPO with intermediate translation-quality rewards to reduce latency without word-level alignment heuristics. Without aligned data in the title means without word-level alignment, not absence of sentence-aligned data or supervision. Inherits Hibiki/Moshi and GRPO. Quality/latency claims are finite multilingual experiments, not universal translation guarantees.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Without aligned data in the title means without word-level alignment, not absence of sentence-aligned data or supervision. Inherits Hibiki/Moshi and GRPO. Quality/latency claims are finite multilingual experiments, not universal translation guarantees.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Simultaneous Speech-to-Speech Translation Without Aligned Data
• Kyutai scientific publications index

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction