Kyutai Releases Voice of Reason Speech-Native Math Model

Kyutai has released Voice of Reason, a pair of open-weight speech-to-speech models designed to solve mathematical problems spoken aloud. These models operate without an intermediate transcription step or a separate text-based large language model (LLM) in the processing loop, directly converting spoken input into spoken mathematical reasoning and solutions. Both models are built upon the GLM-4-Voice-9B architecture and incorporate supervised fine-tuning (SFT) and reinforcement learning (RL) techniques. The direct speech-to-speech approach aims to overcome limitations of cascaded pipelines, which involve speech-to-text, text LLM, and text-to-speech stages. While cascaded systems currently lead in reasoning accuracy, each stage introduces latency and can result in the loss of crucial paralinguistic cues such as tone of voice. Speech-native models, to maintain interactivity, must generate audio output at regular intervals, which constrains the number of internal reasoning tokens they can process. The base GLM-4-Voice model achieved an accuracy of 27.3% on the GSM8K spoken math benchmark. A previous method, STITCH, improved this to 58.7% by incorporating reasoning chunks. Kyutai's research team claims their work represents the first application of reinforcement learning to mathematical reasoning within speech-native AI models. The training process for Voice of Reason involves two stages. The first stage, supervised fine-tuning (SFT), utilized 150,616 problems sourced from Orca-Math. Each problem was rewritten for speech by Qwen3-235B and then vocalized in various voices using Kyutai's DSM TTS system. This SFT stage alone boosted the GLM-4-Voice model's accuracy from 27.3% to 61.7%. The second stage employs reinforcement learning (RL). For each spoken question, the model generates four potential responses at a temperature of 0.9. A separate judge model, Qwen3-235B-A22B-2507, evaluates the decoded text of these responses with a binary reward, without access to the correct answer. This judge model demonstrated an 88% agreement rate with human evaluations on 100 hand-checked cases. The rewards are centered within each group, forming a group-relative REINFORCE objective, which is similar to GRPO but omits PPO clipping and KL regularization. The training was conducted on 16 H100 GPUs, involving approximately 1,500 RL updates. Voice of Reason achieves a significant accuracy of 77.1% on spoken GSM8K, a substantial improvement over the base model's 27.3%. The models are available for self-hosting, with both BF16 checkpoints having been run on a single H100 GPU. Users also require the GLM-4-Voice repository for its speech tokenizer and decoder. The weights are distributed under the GLM-4-Voice license, and as of the release, no inference providers on Hugging Face host these models. The GLM-4-Voice model interleaves its output, producing 13 text tokens followed by 26 audio tokens in a repeating pattern.