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

Kyutai released Voice of Reason, two open-weight speech-to-speech models based on GLM-4-Voice-9B that solve math problems aloud using supervised fine-tuning and reinforcement learning, achieving 77.1% accuracy on spoken GSM8K. The models run efficiently on a single H100 GPU without transcription or separate text LLMs, marking the first use of RL for math reasoning in speech-native models.