Talk to https://t.co/CpQTspHXbi 🔊, the most modular voice AI around. Empower any text LLM with voice, instantly, by wrapping it with our new speech-to-text and text-to-speech. Any personality, any voice. Interruptible, smart turn-taking. We’ll open-source everything within the

🚀New models: ARC-Encoders We introduce a lightweight encoder that compresses context into continuous representations for LLMs, reducing inference cost while preserving performance. Our Adaptable text Representations Compressor, named ARC-Encoder, achieves large efficiency gains

When combined, our TTS and STT can provide a voice to any text LLM, VLM, etc., allowing real-time voice interaction. Try that out in our demo: https://t.co/CpQTspIv0Q Paper: https://t.co/ZvsEElbAmE Models: https://t.co/KjDMx7Kgdd TTS evaluation:

We also train a model on a mix of publicly available datasets, to show that the improvements come from the architecture and not just proprietary data.

We train decoder-only models on time-aligned text and audio data, delaying the “output stream” so that it can be predicted from the “input stream”. At inference, we teacher-force the input stream and predict the output stream. This way, we can build both TTS and STT.

The throughput gains come from the fact that DSMs are easy to batch. The same is true for Kyutai Speech-To-Text, which outperforms Whisper-Streaming in throughput by two orders of magnitude.

We’ve released a preprint on Delayed Streams Modeling (DSM), the framework behind our open, streaming text-to-speech and speech-to-text. Kyutai TTS, powered by DSM, is blazingly fast and competitive with SotA models in quality while providing the best voice cloning. 🧵

🗣️We can listen and speak simultaneously when we talk, and so should the spoken dialogue models (SDMs)! 💬Unlike typical "walkie-talkie" voice AIs, full-duplex SDMs let both sides talk at once - more like real, natural conversation. But this makes alignment harder: - No

If you're at #ICML2025 this week, come check out these 3 posters from our lab🟢! - Aligning Spoken Dialog Models from User Interactions, @anne_youw Thu 17 Jul 11am-1:30pm W-316 - High-Fidelity Simultaneous Speech-To-Speech Translation, @tom_labiausse Wed 16 Jul 4:30pm - 7pm

I’m happy to share that I’ll be attending ICML 2025 in Vancouver next week to present 𝐇𝐢𝐛𝐢𝐤𝐢 [ https://t.co/q50LBOvM1h] 🇫🇷🇬🇧 — Kyutai’s real-time and expressive speech translation system. I'll be presenting the poster on Wednesday, July 16 at 4:30PM, feel free to stop by! 💬

Kyutai TTS comes with hundreds of voices based on Expresso and VCTK. If you would like to see more voices, help us by donating your voice at

Unmute turns a text LLM into a voice AI. At https://t.co/CpQTspIv0Q, it’s @MistralAI's Mistral-Small-3.2-24B, making it fully open-source. Play a quiz game with a snarky host, catch up on tech news, or just hang out and talk. Or modify it to do anything you want!

We’re also releasing the code for https://t.co/CpQTspIv0Q, the modular voice AI system. Make your own personal assistant, make it role-play, give a voice to your agent, or connect it to external tools. You can make it fit onto a single GPU. https://t.co/TgNPIBkuDD

Kyutai TTS and Unmute are now open source! The text-to-speech is natural, customizable, and fast: it can serve 32 users with a 350ms latency on a single L40S. Try it out and get started on the project page: https://t.co/B4P9FuOrQc

Available in PyTorch, MLX, on your iPhone, or in Rust for your server needs! Project Page: https://t.co/bQMP56XIAa OpenASR Leaderboard:

Our latest open-source speech-to-text model just claimed 1st place among streaming models and 5th place overall on the OpenASR leaderboard 🥇🎙️ While all other models need the whole audio, ours delivers top-tier accuracy on streaming content. Open, fast, and ready for production!

The other model is a lightweight English/French 1B model optimized for real-time voice chat apps like https://t.co/CpQTspHXbi. It comes with a semantic voice activity detector that predicts if you’re done talking or just pausing mid-sentence. The open-source releases of Kyutai

Today we are releasing two models. The first one is a 2.6B English-only model that beats Whisper Large v3 on benchmarks even though it’s a streaming model that doesn’t process all the audio at once. It can process 400 sequences in parallel on a single H100.

Kyutai Speech-To-Text is now open-source! It’s streaming, supports batched inference, and runs blazingly fast: perfect for interactive applications. Check out the details here: https://t.co/bQMP56XaKC

Unmute meets Moshi 🫂💖 Talk to https://t.co/CpQTspHXbi!

Using Unmute with a custom voice and prompt to create a very intense ice cream seller, inspired by Justin Kuritzkes' sketch🍦