Voxtral Realtime: causal encoding and adaptive delay conditioning

What changed

Extended DelayedStreamsModeling with a causal audio encoder and adaptive RMSNorm delay conditioning, training a multilingual recognizer that supports variable alignment delays.

Scientific scope

Extended DelayedStreamsModeling with a causal audio encoder and adaptive RMSNorm delay conditioning, training a multilingual recognizer that supports variable alignment delays. Explicitly inherits DSM and language-model components; not invention of streaming ASR. Delay-conditioned benchmark parity at selected settings is not latency/quality equality on every domain. Alignment/pretraining costs remain; accepts speech and outputs text, unlike VoxtralTTS.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Explicitly inherits DSM and language-model components; not invention of streaming ASR. Delay-conditioned benchmark parity at selected settings is not latency/quality equality on every domain. Alignment/pretraining costs remain; accepts speech and outputs text, unlike VoxtralTTS.

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.

• Voxtral Realtime
• Mistral research 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