Kyutai launches voice of reason: a speech-model that solves math problems verbally using reinforcement learning

Kyutai has launched Voice of Reason, a speech-native model that solves math problems through spoken reasoning, utilizing reinforcement learning. The model, built on GLM-4-Voice-9B, incorporates supervised fine-tuning and reinforcement learning without requiring transcription or a separate text-based LLM. It achieves 77.1% accuracy on spoken GSM8K, up from 27.3% in the base model. Deployable for self-hosting, the model runs on a single H100 GPU with BF16 checkpoints, though it requires the GLM-4-Voice repo for speech processing. Kyutai claims this is the first application of RL to math reasoning in speech-native models, with training conducted on 16 H100 GPUs and 1,500 RL updates. The model uses a group-relative REINFORCE objective, similar to GRPO but without PPO clipping or KL regularization. The released checkpoints reflect top-k 50 decoding, with results averaging 77.1% across three seeds. The model’s output interleaves text and audio tokens, with 13 text tokens followed by 26 audio tokens. Evaluation audio came from GPT-4o-mini-TTS, distinct from the training audio. Kyutai also highlights the limitations of cascaded pipelines, which introduce latency and lose paralinguistic cues. The project’s paper, model, and Stitch model are available for further exploration, with credit given to the researchers involved.