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

Kyutai has released Voice of Reason, a pair of open-weight, speech-to-speech models derived from GLM-4-Voice-9B. By utilizing reinforcement learning and supervised fine-tuning, the models achieve a 77.1% accuracy rate on spoken GSM8K math problems, a significant improvement over their baseline performance. Because the system performs reasoning without an intermediate text transcription step or a separate text-based language model, it demonstrates a more efficient architecture for direct voice-based problem solving.