AI
This Model Solves Math Problems Out Loud, and It Is Open for Everyone
Kyutai, the Paris AI lab, just released Voice of Reason, a pair of open weight speech to speech models that solve math problems entirely out loud. Speech goes in, reasoning happens in the audio domain, and speech comes out, with the text detour removed from the pipeline.
The numbers are striking. Starting from the GLM-4-Voice 9B base model, which scored 27.3 percent on spoken GSM8K math problems, Kyutai applied supervised fine tuning on more than 150,000 math problems voiced in many synthetic voices, then added reinforcement learning. The result was 77.1 percent accuracy on spoken math for the full model, and 70.3 percent for a direct variant that answers with zero extra reasoning tokens.
The reinforcement learning step is the research headline. Kyutai describes it as the first application of RL to math reasoning in speech native models, and the training details show how different this discipline is from text. A judge model scored sampled replies with binary rewards, and two design choices proved decisive. Temperature correction, dividing logits by the sampling temperature before the loss, was essential. Removing it collapsed accuracy from 65.5 percent to 12.3 percent. Audio token merging, which asks the loss only whether audio came next rather than which audio token, kept the estimator unbiased with lower variance.
The team also shipped a STITCH variant that writes silent 100 token reasoning chunks between spoken blocks. Because later chunks generate while earlier speech plays, the thinking adds zero extra latency, a neat trick for keeping voice interaction fluid.
Both 9B checkpoints are open and run on a single H100, and the gains survived real speech tests. Transcribed with a separate speech recognizer, the STITCH model held at 72 percent, and voice naturalness scores barely moved after RL. RL even shortened answers, cutting the direct model's average reply from 41.9 to 36.4 seconds.
The training economics carry their own lesson. RL helped most where data was scarcest. With only 10 percent of the fine tuning data, a longer RL run reached 58.5 percent, well above the 43.9 percent that supervised tuning alone managed. The team also ran contamination checks, dropping four overlapping benchmarks and verifying paraphrase overlap against the training set. Open weights, honest evals, and a recipe others can rerun. This is the release template the field should copy.
Why does this matter beyond the benchmark? Voice assistants are moving into ears, cars, and factory floors, and cascade pipelines that hop between speech recognition, text models, and speech synthesis add latency and strip out tone. A model that thinks in sound points to assistants that reason the way humans do in conversation. Kyutai just handed the whole field the recipe, weights included, and that is how open research compounds.
Quick answers
What is this story about?
Kyutai, the Paris AI lab, just released Voice of Reason, a pair of open weight speech to speech models that solve math problems entirely out loud. Speech goes in, reasoning happens in the audio domain, and speech comes out, with the text detour removed from the pipeline.
Why does this story matter?
Why does this matter beyond the benchmark? Voice assistants are moving into ears, cars, and factory floors, and cascade pipelines that hop between speech recognition, text models, and speech synthesis add latency and strip out tone. A model that thinks in sound points to assistants that reason the way humans do in conversation. Kyutai just handed the whole field the recipe, weights included, and that is how open research compounds.
Sources
- MarkTechPost: Kyutai Voice of Reason technical breakdown
- Radoxo AI News Flash: September 23 AI roundup
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.