Research Scientist @ Mistral
Paris, France
Joined September 2021
- Tweets33
- Following209
- Followers107
- Likes175
Pinned Tweet
Language modeling is ambiguous. If you ask someone to caption an image, describe an audio clip, or to translate a sentence, you’ll get several valid outputs.
In this work, we handle this with a winner-takes-all training objective that encourages diverse, plausible predictions.
1/ Multiple Choice Learning of Low-Rank Adapters for Language Modeling
TL;DR: LoRA heads + Winner-Takes-All loss → diverse, plausible LLM outputs
📄 arxiv.org/abs/2507.10419
💻 github.com/Victorletzelter/L…
@VLetzelter, H. Malard, M. Fontaine, G. Richard, S. Essid, @abursuc, @ptrkprz
Victor Letzelter @ ICML 2026 retweeted
Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing.
It detects when someone has finished speaking so agents can respond without delay.
We are releasing new speech models! Happy to have contributed with the team - mostly to the post-training of the Voxtral 3 transcribe models - which are quite robust and with diverse capabilities (offline and realtime multilingual transcription, diarization, context bias)
Mistral has released new Speech models! Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard, and Voxtral 3 TTS enters the top 12 in both Text to Speech arenas
@MistralAI has released Voxtral 3: a new Text to Speech model and three Speech to Text models (Small Transcribe 3, Mini Transcribe 3 and Mini Transcribe Realtime 3). We evaluated all of them.
Key takeaways:
➤ Speech to Text: Voxtral Small Transcribe 3 takes the #3 spot on the Artificial Analysis Speech to Text leaderboard with a 2.1% AA-WER Index, ahead of ElevenLabs Scribe v2 (2.2%) and Grok Voice Transcribe 2.0 (2.3%). Voxtral Mini Transcribe 3 lands at #5 with 2.2%.
➤ Streaming Speech to Text: Voxtral Mini Transcribe Realtime 3 scores 3.3% final-transcript AA-WER at the 960 ms setting (#6 of 46) and 3.6% at 480 ms, down from 5.2% for the previous Voxtral Mini Transcribe Realtime.
➤ Text to Speech, Provider Voice: Voxtral 3 TTS scores an Elo of 1,218 across 1,599 arena appearances, ranking #12 of 98, just behind Breeze TTS 2 at 1,221 and ahead of Inworld Realtime TTS-2 Flash and StepAudio 2.5 TTS at 1,216. It ranks #13 in English (US) and #12 in English (UK).
➤ Text to Speech, Controlled Voice: Voxtral 3 TTS scores 1,068 across 1,614 appearances, ranking #12 of 43 with the same 8 cloned voices, behind Eleven v3 at 1,072 and level with Ant Group's Ming-3.0-tts at 1,068. It ranks #10 in Assistants and #10 in English (UK).
➤ Pronunciation Robustness: Voxtral 3 TTS scores 84.4%, ranking #13 of 32 models, behind Qwen-Audio-3.1-TTS-Plus at 84.6% and ahead of HeyGen Voice at 83.2%. It ranks #3 in Standalone Terms at 93.9%.
➤ Pricing: Voxtral 3 TTS is priced at $13/1M characters, compared to Gemini 3.8 Flash TTS at $16.5/1M characters, Eleven v4 Turbo at $40/1M characters and Simba 3.2 at $6.6/1M characters.
See more details and listen to samples below ⬇️
Read the blog post here mistral.ai/news/voxtral-3/, more content to come!
Victor Letzelter @ ICML 2026 retweeted
1/n Our paper REPA-G is accepted at #NeurIPS2026! 🎉 I'll present it at the NeurIPS Paris site 🇫🇷.
We introduce test-time conditioning of REPA-style diffusion models with visual features, with no extra training.
Paper: arxiv.org/abs/2507.10419
Code: github.com/Victorletzelter/L…
Video: icml.cc/virtual/2026/poster/…
Victor Letzelter @ ICML 2026 retweeted
Introducing Mistral OCR 4. It creates structure with bounding boxes, block classification, and inline confidence scores in 170 languages. 🧵👇
Victor Letzelter @ ICML 2026 retweeted
Congrats to our @VLetzelter for earning his well deserved PhD with his works on "Multiple Choice Learning from Ambiguous Signals".
Victor is a resourceful, curious and kind researcher. Good luck for next!
Fruit of an excellent collaboration between @telecomparis & @valeoai
🚨 Happening NOW! 🎓
Our PhD student Victor Letzelter @VLetzelter is currently defending his thesis at Telecom Paris!
Good luck, Victor! 👏🚀 #PhDDefense #TelecomParis
Victor Letzelter @ ICML 2026 retweeted
1/ 📢 New preprint: Test-Time Conditioning with Representation-Aligned Visual Features
Introducing REPA-G — a framework for controllable image generation at inference time using aligned visual features.
📄 arxiv.org/pdf/2602.03753
💻 github.com/valeoai/REPA-G
🧵👇
Let’s catch up if you’re at @icmlconf and want to chat!
Interested in time series forecasting and data uncertainty quantification?
Check out our latest paper with @VLetzelter at @icmlconf !
Paper: arxiv.org/abs/2506.05515
Code: github.com/Victorletzelter/t…
Poster #2211 , Tue 15 Jul 11 a.m. PDT East
#timeseries #quantization #uncertainty
Victor Letzelter @ ICML 2026 retweeted
Inferring 3D human poses from video is highly ill-posed because of depth ambiguity.
Our work accepted to #NeurIPS2024, ManiPose, gets one step closer to solving this, by leveraging prior knowledge about poses topology and cool multiple-choice learning techniques.
Victor Letzelter @ ICML 2026 retweeted
🌟 Calling all MSc students passionate about computer vision and ML!
We’re offering research internships about diffusion models, multi-modal transformers, continual learning, & more. 4 exciting openings await!
🔗 Learn more: valeoai.github.io/interns/
RT to spread the word! 🙌
Working on ill-posed machine learning tasks, interested in multi-heads neural networks and data #uncertainty quantification ?
Sharing here our latest research, which will be presented at @NeurIPSConf in December.
Key insights:
* Annealing enhances exploration compared to greedy convergence.
* Inspired by statistical physics and information theory, we describe the training trajectory.
* Experiments on synthetic datasets, UCI benchmarks, and speech separation show highly promising results.
Check out the paper!
Arxiv: arxiv.org/abs/2407.15580
Project page: valeoai.github.io/publicatio…
GitHub: github.com/Victorletzelter/a…
A joint work with David Perera, Théo Mariotte, Adrien Cortes, @Mickael_Chen , Slim Essid and Gaël Richard at @valeoai and Télécom Paris.