Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning
https://www.marktechpost.com/2026/09/22/kyutai-releases-voice-of-reason-a-speech-native-model-that-solves-spoken-math-with-reinforcement-learning/📌 語音直接推理數學：Kyutai 用 RL 把準確率從 27% 推到 77%
TL;DR：Kyutai 開權重發布 Voice of Reason，語音原生模型不經轉文字，靠 RL 大幅提升口語數學能力。
想像你直接對著 AI 唸一道數學題，它不轉文字、不呼叫另一個文字 LLM，就用聲音直接回答你——而且準確率能從不到三成推到接近八成。
🤔 背景：語音原生模型的兩難
摘要指出，cascaded pipeline（語音轉文字 → 文字 LLM → 文字轉語音）在推理任務上目前仍然領先，但每一段都會疊加延遲，也會丟失語氣等 paralinguistic 線索。語音原生模型則必須定時吐出音訊才能維持互動感，這限制了它能負擔多少「隱藏推理」token。
🧩 方法與架構：13 個文字 token 換 26 個音訊 token
Kyutai 發布的兩個開權重 speech-to-speech 模型，命名為 Voice of Reason，都以 GLM-4-Voice-9B 為起點，接著疊加 supervised fine-tuning（SFT）與強化學習（RL），整個過程沒有轉錄步驟，也沒有另外接一個文字 LLM。GLM-4-Voice 原本的輸出方式是交錯進行：13 個文字 token，接著 26 個音訊 token，反覆循環。
訓練用的 RL 目標，把 reward 在每個 group 內置中，形成一個 group-relative REINFORCE 目標，研究團隊表示這與 GRPO 有關，但拿掉了 PPO 的 clipping 與 KL regularization。訓練用了 16 張 H100，總共跑了 1,500 次 RL 更新。
📊 數據：27.3% → 58.7% → 77.1%
- Base GLM-4-Voice 在口語版 GSM8K 上的準確率是 27.3%。
- 先前的 STITCH 方法透過加入推理片段，把它推到 58.7%。
- 論文報告的分數採用 top-k 50 decoding、3 個 seed 平均；若拿掉 top-k 限制，兩個發布的 checkpoint 分別達到 70.3% 與 77.1%。
- 評估用的語音來自 GPT-4o-mini-TTS，與訓練資料使用的 TTS 系統不同，判分則由 GPT-4o 擔任 judge。
- 研究團隊稱這是 RL 首次應用在語音原生模型的數學推理任務上。
⚠️ 限制
摘要指出，cascaded pipeline 在純推理分數上仍然領先，只是要用延遲與 paralinguistic 線索換取；文中列出的 omni 與 cascaded 系統分數屬於更大規模的頂線參考，並非同規格的對照比較；評估語音的 TTS 系統與訓練資料的 TTS 系統不同，可能造成分布上的落差。
🎯 實務啟示
兩個 BF16 checkpoint 都可以在單張 H100 上自行架設，但需要另外準備 GLM-4-Voice repo 提供的 speech tokenizer 與 decoder，權重繼承 GLM-4-Voice 的授權條款，目前也沒有 Hugging Face inference provider 託管。對想實驗「語音直接推理」而非傳統 cascaded 架構的團隊來說，這是目前少見的開權重起點。
🔗 來源
- 標題：Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning
- 作者／機構：Asif Razzaq／MarkTechPost
- 連結：https://www.marktechpost.com/2026/09/22/kyutai-releases-voice-of-reason-a-speech-native-model-that-solves-spoken-math-with-reinforcement-learning/
#Kyutai #SpeechAI #ReinforcementLearning #GLM4Voice #VoiceAI #GSM8K #OpenWeights #SpeechToSpeech #MathReasoning #AudioML
原始資料 MarkTechPost · 收集於 2026-09-23
摘要原文
Kyutai has released Voice of Reason , 2 open-weight speech-to-speech models that solve math problems out loud. Both start from GLM-4-Voice-9B and add supervised fine-tuning (SFT) and reinforcement learning (RL). There is no transcription step and no separate text LLM in the loop. On spoken GSM8K, accuracy climbs from 27.3% for the base model to 77.1%. Is it deployable? Yes, for self-hosting. Kyutai ran both BF16 checkpoints on a single H100. You also need the GLM-4-Voice repo for its speech tokenizer and decoder. Weights inherit the GLM-4-Voice license, and no Hugging Face inference provider hosts them yet. Cascaded pipelines (speech-to-text, text LLM, text-to-speech) still lead on reasoning. However, each stage adds latency, and the pipeline loses paralinguistic cues like tone. Speech-native models must emit audio at regular intervals to stay interactive. That limits how many hidden reasoning tokens they can afford. Base GLM-4-Voice scores 27.3% on GSM8K . The earlier STITCH method raised that to 58.7% by adding reasoning chunks. The research team call their work the first application of RL to math reasoning in speech-native models. GLM-4-Voice interleaves its output: 13 text tokens, then 26 audio tokens, repeating. Rewards are centered within each group, forming a group-relative REINFORCE objective. It is related to GRPO but drops PPO clipping and KL regularization. Training ran on 16 H100 GPUs, with 1,500 RL updates. Paper scores use top-k 50 decoding, averaged over 3 seeds. Removing top-k gives 70.3% and 77.1%. The released checkpoints correspond to these runs. The omni and cascaded systems are larger top lines, not matched comparisons. Evaluation audio came from GPT-4o-mini-TTS, a different TTS system than the training audio. GPT-4o served as the evaluation judge. Check out the Paper , the direct model and the Stitch model . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成