Kyutai released Voice of Reason, two open‑weight speech‑to‑speech models that solve math problems out loud without a separate transcription step or text LLM in the loop. Built on GLM‑4‑Voice‑9B, they add supervised fine‑tuning on 150 k spoken math problems and a reinforcement‑learning stage that samples multiple answers, scores them with an LLM judge, and updates the policy using a group‑relative REINFORCE objective. On spoken GSM8K accuracy rises from 27.3 % for the base model to 77.1 % for the Stitch variant.
The main pain points for developers are latency, loss of paralinguistic cues, and limited hidden reasoning tokens when models must emit audio at regular intervals. Cascaded pipelines (speech‑to‑text → text LLM → text‑to‑speech) still lead on pure reasoning but add delay and strip tone. Voice of Reason avoids these drawbacks by keeping the entire reasoning process in the speech domain.
Deployment is straightforward for self‑hosting. The BF16 checkpoints run on a single H100 GPU. You need the GLM‑4‑Voice repository for its speech tokenizer and decoder; the model weights inherit the GLM‑4‑Voice license and are not yet hosted on any Hugging Face inference provider. No extra inference server is required beyond loading the checkpoints and using the provided tokenizer/decoder.
Two design choices proved critical: temperature correction during RL loss computation prevents collapse of performance, and merging audio‑token probabilities into an abstract token yields an unbiased, lower‑variance gradient estimator. Together they enable stable training and the observed accuracy gains.
For teams wanting real‑time, interactive math tutoring or voice‑first educational tools, Voice of Reason offers a practical path: a single‑GPU deployable model that reasons aloud, retains expressive speech cues, and avoids the latency and complexity of cascaded systems. Test the checkpoints, adapt the SFT+RL pipeline to your domain, and integrate the tokenizer/decoder into your voice‑assistant stack.
#AI #ML #SpeechAI #ReinforcementLearning #MathReasoning #DeployAI