Kyutai has released Voice of Reason, 2 open-weight speech-to-speech models that solve math problems out loud. Both start from GLM-4-Voice-9B and add supervised fine-tuning (SFT) and reinforcement learning (RL). There is no transcription step and no separate text LLM in the loop. On spoken GSM8K, accuracy climbs from 27.3% for the base model to 77.1%.

Is it deployable? Yes, for self-hosting. Kyutai ran both BF16 checkpoints on a single H100. You also need the GLM-4-Voice repo for its speech tokenizer and decoder. Weights inherit the GLM-4-Voice license, and no Hugging Face inference provider hosts them yet.

Why Speech Models Lag on Math

Cascaded pipelines (speech-to-text, text LLM, text-to-speech) still lead on reasoning. However, each stage adds latency, and the pipeline loses paralinguistic cues like tone. Speech-native models must emit audio at regular intervals to stay interactive. That limits how many hidden reasoning tokens they can afford.

Base GLM-4-Voice scores 27.3% on GSM8K. The earlier STITCH method raised that to 58.7% by adding reasoning chunks. The research team call their work the first application of RL to math reasoning in speech-native models.

How the Training Works

GLM-4-Voice interleaves its output: 13 text tokens, then 26 audio tokens, repeating.

• Stage-1 SFT: Training uses 150,616 problems from Orca-Math. Qwen3-235B rewrote each problem for speech. Kyutai’s DSM TTS then voiced them in many voices. SFT alone lifts GLM-4-Voice from 27.3% to 61.7%.
• Stage-2 RL: For each spoken question, the model samples 4 replies at temperature 0.9. A judge, Qwen3-235B-A22B-2507, scores the decoded text with a binary reward. The judge 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. It is related to GRPO but drops PPO clipping and KL regularization. Training ran on 16 H100 GPUs, with 1,500 RL updates.

2 design choices matter most:

• Temperature correction: Logits are divided by the sampling temperature before the log-softmax in the loss. Without it, GSM8K collapsed from 65.5% to 12.3%.
• Audio-token merging: At each audio position, all audio-vocabulary probabilities are summed into 1 abstract token. 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.

Interactive Explainer

&&&

The 2 Released Checkpoints

• glm-4-voice-of-reason-9b answers directly, with no extra reasoning tokens. Any step-by-step working is spoken aloud.
• glm-4-voice-of-reason-stitch-9b 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.

Results

Model	Params	GSM8K (%)
PersonaPlex (full-duplex)	8B	3.2
GLM-4-Voice	9B	27.3
STITCH (Chiang et al.)	9B	58.7
Voice of Reason	9B	65.5 ± 1.1 (70.3 released)
Voice of Reason (Stitch)	9B	74.8 ± 1.1 (77.1 released)
Qwen2.5-Omni (text output)	7B	84.7
Qwen3-Omni (text output)	30B	94.6
Cascaded ASR-LLM-TTS-ASR	31B LLM	95.7

Paper scores use top-k 50 decoding, averaged over 3 seeds. Removing top-k gives 70.3% and 77.1%. The released checkpoints correspond to these runs. The omni and cascaded systems are larger top lines, not matched comparisons.

Other findings:

• Gains survive in real speech: Transcribed with Qwen3-ASR-1.7B, the Stitch model scores 72.0 ± 1.9%.
• Naturalness holds: UTMOSv2 moved from 4.067 to 4.069 (direct) and 4.174 to 4.164 (Stitch) after RL.
• Gains do not come from longer answers: RL cut the direct model’s average reply from 41.9 to 36.4 seconds. 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%. SFT alone scored 43.9%.
• General knowledge dips: Spoken TriviaQA fell from 40.6% to 34.0% for the direct model. The authors attribute this mainly to full-data SFT, not RL.
• 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. GPT-4o served as the evaluation judge.

Key Takeaways

• Kyutai’s RL recipe lifts GLM-4-Voice from 27.3% to 77.1% on spoken GSM8K.
• The direct model reaches 70.3% without reasoning tokens, beating STITCH’s 58.7%.
• Temperature correction is critical: removing it crashed accuracy to 12.3%.
• Speech naturalness holds after RL; TriviaQA drops, mostly from SFT.
• Both 9B checkpoints are open on Hugging Face and run on 1 H100.

---

Check out the Paper, the direct model and the Stitch model. All credit goes to the researcher of this project. Also, feel free to follow us on