Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning

Spoken interfaces still trail cascaded ASR–LLM–TTS pipelines on hard reasoning. MarkTechPost covers Kyutai’s September 22, 2026 release of Voice of Reason: two open-weight 9B speech-to-speech models based on GLM-4-Voice that solve math out loud via supervised fine-tuning plus RL — no transcription step, no separate text model mid-flight.

On spoken GSM8K, the Stitch checkpoint reaches 77.1% (released, without top-k) versus 27.3% for the base model and 58.7% for the earlier STITCH method. Both BF16 checkpoints run on a single H100; weights inherit the GLM-4-Voice license and need that repo’s speech tokenizer and decoder.

We’re covering this as a bet that voice-native reasoning can close the gap without surrendering interactivity.

Cascades still win the top line — Qwen3-Omni text output hits 94.6%, and a 31B cascaded stack 95.7% — but every stage adds latency and strips tone. Speech-native models must emit audio on a schedule; that caps hidden reasoning tokens. Kyutai’s recipe: Stage-1 SFT on 150,616 Orca-Math problems rewritten for speech by Qwen3-235B and voiced via Kyutai’s DSM TTS, lifting accuracy to 61.7%. Stage-2 RL samples four replies per question; Qwen3-235B-A22B-2507 judges decoded text with a binary reward (88% human agreement on 100 hand checks), using a group-relative REINFORCE objective related to GRPO without PPO clipping or KL.

Two design choices are the story under the scoreboard. Temperature correction — dividing logits by sampling temperature before log-softmax — prevented a collapse from 65.5% to 12.3%. Audio-token merging treats all audio-vocab mass as one abstract token so the loss only asks whether audio came next; the paper claims an unbiased, lower-variance estimator under value invariance. The direct model speaks its working; the Stitch model writes silent 100-token reasoning chunks that can generate while earlier speech plays, adding no extra latency in Kyutai’s framing.

Useful AI loves open speech models that get better at real tasks people actually speak — tutoring, accessibility, hands-busy work. Naturalness held (UTMOSv2 barely moved). Gains survive ASR re-transcription (~72% for Stitch). RL even shortened average direct replies. The harm and caveat side: spoken TriviaQA fell from 40.6% to 34.0% (authors blame mostly full-data SFT); paraphrase overlap with Orca-Math hit 54.0% of 678 checked questions after dropping some contaminated sets; omni/cascaded systems still lead; no Hugging Face inference provider hosts them yet.

I’m watching whether the RL recipe transfers beyond GSM8K-style math, whether contamination controls tighten, and whether voice-native models start eating cascaded latency budgets in products.

Coverage by Asif Razzaq at MarkTechPost. Released checkpoints: glm-4-voice-of-reason-9b (direct) and glm-4-voice-of-reason-stitch-9b. Training used 16 H100s and 1,500 RL updates. Eval audio came from GPT-4o-mini-TTS, a different TTS than training.