Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning
MarkTechPost Asif Razzaq
Kyutai released two open-weight speech models that solve math out loud. The big twist: one gets 77.1% on spoken GSM8K without a text LLM in the middle.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Kyutai has put out Voice of Reason, a pair of open-weight speech-to-speech models that can work through math problems aloud. They start from GLM-4-Voice-9B, then add supervised fine-tuning and reinforcement learning. There’s no transcription stage and no separate text model sitting in the middle, which is exactly the kind of shortcut speech systems have been avoiding for years.
The headline number is striking. On spoken GSM8K, the base model scores 27.3%. After the training recipe, that rises to 77.1% for the released Stitch checkpoint. Kyutai also says the direct version reaches 70.3% without extra reasoning tokens, which already beats the earlier STITCH result of 58.7% from the same paper family. For a field where cascaded speech-to-text-to-text-to-speech setups still dominate reasoning, that is a meaningful jump.
The training setup is unusual in a useful way. Stage one uses 150,616 Orca-Math problems rewritten for speech by Qwen3-235B and voiced by Kyutai’s DSM TTS in many voices. That alone lifts GLM-4-Voice to 61.7%. Stage two adds reinforcement learning: for each spoken question, the model samples four replies at temperature 0.9, then a judge model scores the decoded text with a binary reward. Kyutai says the judge agreed with humans 88% of the time on 100 hand-checked cases.
Two details seem to carry a lot of the weight. First, temperature correction: if the logits are not divided by the sampling temperature before the loss, GSM8K drops from 65.5% to 12.3%. Second, audio-token merging: the loss only asks whether audio came next, not which audio token, by summing all audio-vocabulary probabilities into one abstract token. Kyutai argues that makes the estimator unbiased and lower-variance under a value-invariance assumption.
There are tradeoffs, because there always are. Speech naturalness stays roughly intact after RL, but spoken TriviaQA falls from 40.6% to 34.0% in the direct model, which the authors mostly blame on full-data SFT. The good news for actual users is that both BF16 checkpoints can run on a single H100, though you still need the GLM-4-Voice repo for the speech tokenizer and decoder, and no Hugging Face inference provider hosts them yet.
My take — AI-written commentary, not fact-checked reporting
This is the kind of result that makes the “just run a text model behind the scenes” crowd look a bit lazy. Speech-native reasoning is still awkward, but Kyutai showed it can do real work without hiding behind a cascaded pipeline. The more interesting part is that the gains came from training discipline, not from sprinkling hype on audio and calling it intelligence.
Read more about this at: MarkTechPost