Alexandre Défossez

Learn more about it with Vaclav, lead engineer for Unmute, getting interviewed by the AI:

We just released unmute.sh 🔇🔊
It is a text LLM wrapper, based on in-house streaming ASR, TTS, semantic VAD to reduce latency. ⏱️
Unlike Moshi 🟢, Unmute 🔊 is turn base, but allows customization in two clicks🖱️: voice and prompt!
Paper and open source coming soon.

We just open sourced a fine tuning codebase for Moshi!

Just back from holidays, so a bit late, to announce MoshiVis, extending Moshi's multimodal capabilities to take in images 📷.
Only 200M weights were added to plug a ViT through cross attention with gating 🖼️🔀🎤
Training relies on a mix of text only and text+audio synthetic data (~20k hours) 💽

I'll start my presentation in 10 minutes, you can join in Zoom: concordia-ca.zoom.us/j/81541793947
See you there!

I'll present a dive into Moshi 🟢 and our translation model Hibiki 🇫🇷♻️🇬🇧 as part of the next @convai-rg.bsky.social reading group 👨‍🏫📗.

📅 13th of March 🕰️ 11am ET, 4pm in Paris.

I'll discuss Mimi 🗜️ and multi-stream audio modeling 🔊.
Join on Zoom, replay on YT.

⬛ ⬛ 🟧 🟧 🟨 🟨 🟩 🟩 🟩 ⬛
⬛ 🟧 🟧 🟨 🟨 🟩 🟩 🟩 ⬛ ⬛

Even Kavinsky 🎧🪩 can't break Hibiki! Just like Moshi, Hibiki is robust to extreme background conditions 💥🔊.

France TV - Replay et Direct tv des chaînes France Télévisions (ex Pluzz) Retrouvez toutes les vidéos, articles et podcasts des programmes des chaînes de France Télévisions.

Very happy to have participated in this *beautiful* documentary from Florent Muller, on the frontiers between humans and machines,
following next @yann-lecun.bsky.social and so many humbling figures of AI:
www.france.tv/documentaire...

Our latest studies on the decoding text from brain activity, reviewed by MIT Tech Review @technologyreview.com

www.technologyreview.com/2025/02/07/1...

Building Bridges between Regression, Clustering, and Classification Regression, the task of predicting a continuous scalar target y based on some features x is one of the most fundamental tasks in machine learning and statistics. It has been observed and...

Check out our paper, with Lawrence Stewart and @bachfrancis.bsky.social

Link: arxiv.org/abs/2502.02996

1/8

Excited to meet and exchange with a number of actors from all around the world at the AI Summit 🌍

We just released Hibiki, a 🎙️-to-🔊 simultaneous translation model 🇫🇷🇬🇧
We leverage a large synthetic corpus synthesized from the text translation model MADLAD, and our own TTS + simple lag rule.
Model is decoder only, runs at scale, even on device 📲
github.com/kyutai-labs/hibiki

🚨Job alert (Please RT)

What: masters internship and/or PhD positions
Where: Rothschild Foundation Hospital (Paris, France)
Topic: AI and Neuroscience
Supervised by: Pierre Bourdillon and myself
Apply here: forms.gle/KKnea2QAjhAe...
Deadline: Feb 5th

We just released the Helium-1 model , a 2B multi-lingual LLM which @exgrv.bsky.social and @lmazare.bsky.social have been crafting for us! Best model so far under 2.17B params on multi-lingual benchmarks 🇬🇧🇮🇹🇪🇸🇵🇹🇫🇷🇩🇪
On HF, under CC-BY licence: huggingface.co/kyutai/heliu...