Kyutai's Voice of Reason: Speech Models Solve Spoken Math Without a Text LLM
Kyutai has released 'Voice of Reason', two 9B checkpoints built on GLM‑4‑Voice‑9B that take a spoken math problem and answer in speech, with no transcription step and no separate text LLM in the loop.
Key points from Kyutai’s announcement and paper:
- Base vs. fine‑tuned: On spoken GSM8K, the base GLM‑4‑Voice model scores 27.3%. After supervised fine‑tuning + reinforcement learning, the Voice of Reason models reach ~70.3% (direct) and ~77.1% (STITCH, with silent reasoning chunks).
- Two checkpoints:
glm-4-voice-of-reason-9b – answers directly, with any step‑by‑step reasoning spoken aloud.
glm-4-voice-of-reason-stitch-9b – interleaves silent reasoning while speaking, so “thinking” doesn’t add extra latency.
- Open weights, self‑hosted: Both models are released on Hugging Face under the GLM‑4‑Voice license and can run in BF16 on a single H100 using the GLM‑4‑Voice codebase for the speech tokenizer/decoder. No third‑party inference provider is hosting them yet.
Why it matters -
Shows that reasoning can be embedded directly in audio tokens, not just via cascaded ASR → text LLM → TTS pipelines. Relevant for voice agents, call‑center assistants, and hands‑free agentic workflows.
This is still a research release focused on math benchmarks, not a general “voice AGI” claim but it’s a solid step toward more capable, lower‑latency speech‑native agents.
#AI #VoiceAI #LLM #AgenticAI #OpenSource #MachineLearning
Kyutai has released Voice of Reason, two open-weight speech-to-speech models built on GLM-4-Voice-9B. They take spoken math problems and answer aloud without transcribing the input or routing through a separate text LLM.
Both models were trained in two stages: supervised fine-tuning on about 150,000 spoken math problems, followed by reinforcement learning that rewards correct answers. On spoken GSM8K, accuracy rose from 27.3% for the base GLM-4-Voice to about 70% for the direct-answer model and 77% for the STITCH variant, which interleaves silent reasoning with speech. These are best-run scores; averages across runs were somewhat lower.
The checkpoints are on Hugging Face and run in BF16 on a single H100, but you also need the GLM-4-Voice codebase for speech input and audio output. No third-party inference provider hosts them yet.
The tradeoff is specialization. General knowledge dropped sharply on a spoken TriviaQA benchmark, and cascaded pipelines still lead on reasoning. This is a research release focused on math, but it shows that reasoning can live inside audio-native models, which is relevant for lower-latency voice agents.