Kyutai publishes speech-to-speech models that answer out loud without transcribing

single source· 1 articles · confidence: medium · first seen 2026-09-23 06:33 UTC

What this means for you

The checkpoints are public and fit on one H100, so this is testable today rather than a paper claim. Nothing to migrate: no hosted endpoint is mentioned, and the 77.1% figure comes with no evaluation date or harness, so score it on your own audio before moving any voice pipeline off transcription.

Kyutai has released Voice of Reason, two open-weight speech-to-speech models (audio in, audio out, with no transcription step and no text model in between) built on GLM-4-Voice-9B. Supervised fine-tuning followed by reinforcement learning — training against a reward signal rather than labelled examples — lifted accuracy on spoken GSM8K, a set of grade-school maths word problems read aloud, from 27.3% to 77.1%. Both checkpoints are on Hugging Face and run on one H100. The source does not give the evaluation date or harness behind the score.

Key facts

  • ·Kyutai released Voice of Reason, two open-weight speech-to-speech models built on GLM-4-Voice-9B. source
  • ·Accuracy on spoken GSM8K rose from 27.3% to 77.1% after supervised fine-tuning and reinforcement learning. source
  • ·The models use no transcription step and no text LLM in the loop. source
  • ·Both checkpoints are on Hugging Face and run on a single H100. source

What the sources say

  • MarkTechPost — Release note on the two checkpoints, the training method and the reported spoken-maths gain.

Sources

The original reporting. Follow these — they did the work.

← the wire