Kyutai Releases Voice of Reason, Speech-Native Math Models With 77.1% Spoken GSM8K Accuracy

The two checkpoints are glm-4-voice-of-reason-9b, which answers directly with no extra reasoning tokens and speaks any step-by-step working aloud, and glm-4-voice-of-reason-stitch-9b, which writes silent 100-token reasoning chunks between spoken blocks. Kyutai says later chunks are generated while earlier speech plays, so thinking adds no extra latency. In the STITCH-R layout used here, the first reasoning chunk precedes the first spoken block. Both BF16 checkpoints run on a single H100 for self-hosting. Users also need the GLM-4-Voice repository for its speech tokenizer and decoder. The weights inherit the GLM-4-Voice license, and no Hugging Face inference provider hosts them yet.

Cascaded pipelines that combine speech-to-text, a text LLM, and text-to-speech still lead on reasoning, but each stage adds latency and the pipeline loses paralinguistic cues such as tone. Speech-native models must emit audio at regular intervals to stay interactive, which limits how many hidden reasoning tokens they can afford. The base GLM-4-Voice scores 27.3% on GSM8K, and the earlier STITCH method raised that to 58.7% by adding reasoning chunks. The research team calls its work the first application of reinforcement learning to math reasoning in speech-native models.

GLM-4-Voice interleaves its output as 13 text tokens followed by 26 audio tokens, repeating. In stage-1 supervised fine-tuning, training used 150,616 problems from Orca-Math. Qwen3-235B rewrote each problem for speech, and Kyutai's DSM TTS voiced them in many voices. SFT alone lifts GLM-4-Voice from 27.3% to 61.7%. In stage-2 reinforcement learning, the model samples four replies at temperature 0.9 for each spoken question. A judge, Qwen3-235B-A22B-2507, scores the decoded text with a binary reward and never sees the reference answer. On 100 hand-checked cases, it agreed with humans 88% of the time. Rewards are centered within each group, forming a group-relative REINFORCE objective related to GRPO but without PPO clipping and KL regularization. Training ran on 16 H100 GPUs with 1,500 RL updates.

Two design choices matter most. In temperature correction, logits are divided by the sampling temperature before the log-softmax in the loss; without this, GSM8K collapsed from 65.5% to 12.3%. In audio-token merging, all audio-vocabulary probabilities at each audio position are summed into one abstract token, so the loss asks only whether audio came next, not which audio token. The paper proves this estimator is unbiased and lower-variance under a value-invariance assumption.

Reported scores use top-k 50 decoding, averaged over three seeds. Voice of Reason scores 65.5% plus or minus 1.1 percentage points, with the released checkpoint at 70.3%. Voice of Reason with Stitch scores 74.8% plus or minus 1.1 points, with the released checkpoint at 77.1%. Removing top-k gives the released figures of 70.3% and 77.1%. The direct model reaches 70.3% without reasoning tokens, beating STITCH's 58.7%. For comparison, Qwen2.5-Omni with text output scores 84.7%, Qwen3-Omni with text output scores 94.6%, and a cascaded ASR-LLM-TTS-ASR system with a 31B LLM scores 95.7%. The source describes the omni and cascaded systems as larger top lines, not matched comparisons.

Gains survive in real speech: transcribed with Qwen3-ASR-1.7B, the Stitch model scores 72.0% plus or minus 1.9 points. Naturalness holds, with UTMOSv2 moving from 4.067 to 4.069 for the direct model and from 4.174 to 4.164 for Stitch after reinforcement learning. The gains do not come from longer answers: RL cut the direct model's average reply from 41.9 to 36.4 seconds, and Stitch reasoning tokens rose only from 167 to 176. RL helps most with little data: with 10% of SFT data, a longer RL run reached 58.5%, while SFT alone scored 43.9%.

General knowledge dips. Spoken TriviaQA fell from 40.6% to 34.0% for the direct model, which the authors attribute mainly to full-data SFT, not RL. In a contamination check, the team dropped AddSub, MultiArith, SingleEQ and SVAMP from evaluation. Of 678 checked questions, 54.0% overlapped with Orca-Math at the paraphrase level. Evaluation audio came from GPT-4o-mini-TTS, a different TTS system than the training audio, and GPT-4o served as the evaluation judge.

Both 9B checkpoints are open on Hugging Face and run on one H100. Kyutai also released the paper, the direct model, and the Stitch model.

Editor's Summary Kyutai's Voice of Reason shows that reinforcement learning can lift a speech-native model from 27.3% to 77.1% on spoken GSM8K while maintaining speech naturalness. The two open-weight 9B checkpoints run on a single H100, though no Hugging Face inference provider hosts them yet and spoken TriviaQA declined. The reported gains rely on a judge model, a specific TTS setup for evaluation, and comparisons that include larger non-matched systems.