Eustache Le Bihan

@paulcx @Julianfmack @huggingface @Tu7uruu @TheOneKloud it does! Check out the transformers doc #non-english-transcription" target="_blank" rel="nofollow noopener">huggingface.co/docs/transform…

@Julianfmack @huggingface @Tu7uruu @eustachelb @TheOneKloud Does it support multi-languages?

Happy to share what I've been working on recently: today we release Cohere Transcribe, a state-of-the-art speech recognition model that beats both commercial and open-source models to land at #1 on the Open ASR Leaderboard!

just `uv pip install -U transformers` and play with it!

Been cooking again for you guys! Cohere ASR model, topping the open ASR leaderboard, is supported day 0 in transformers 🤠

@neilzegh @lmazare 🐐

When @lmazare wants some fresh air from optimizing API speech models, he optimizes on-device ones.

In addition to vLLM, the Hugging Face Transformers library now supports Voxtral Realtime. Thanks to @eustachelb for leading the integration. We’ve been amazed by how quickly the OSS community has shipped implementations of Voxtral Realtime across platforms, backends, and use cases. We expect Transformers support to drive even wider adoption, especially across fine-tuning and quantization libraries. We want to thank Mergen Nachin, Salvatore Sanfilippo, TrevorS, Shreyas Karnik, Awni Hannun, and @limzba for their early contributions. Full list of community integrations: #community-contributions-untested" target="_blank" rel="nofollow noopener">huggingface.co/mistralai/Voxt…

Really grateful to @eustachelb for helping me land this in the transformers library and @Tu7uruu for getting the model up on OpenASR leaderboard.

@antirez @julien_c @1littlecoder @MistralAI indeed! one might even notice that STFT frames are shifted by 40ms (using non-streaming inference as a ref)

@eustachelb @julien_c @1littlecoder @MistralAI You can find all the architectural detains into my voxtral.c GitHub repository if it can help in some way. There were details disseminated among mistral-common and vLLM that I reconstructed, also note that the exact FFT points are crucial for the model to work well.

Yesterday @MistralAI released an open weights transcription model able to work in real time, Voxtral Mini 4B. Today, following the Whisper.cpp lesson, here is a C inference pipeline ready to use as a library, I hope you'll enjoy it: github.com/antirez/voxtra…

@julien_c @antirez @1littlecoder @MistralAI soon in transformers indeed! thinking about a quick thread to detail a bit the architecture since their paper will not come right now

@antirez @1littlecoder @MistralAI @eustachelb (With the help of the mistral team)

really nothing difficult, the vllm implem is unnecessarily complicated (since they don't use a conv cache, they have to provide more context, meaning sending longer overlapping audio chunks). The main complication comes from the STFT computation and where to cut when sending audio chunk

Yes, for the "easy work" part. The model inference was only half-specified via the vLLM nightly + mistral common Python stuff. I needed to find my way with the inference using Codex 5.2 xhigh, only way to understand enough details. Then, with MODEL.md, Claude did it mostly.

The founders of @huggingface after turning down $500m from Nvidia so they can stay independent

cooking for you guys 🧑‍🍳 really really soon in transformers

Today is a big day, transformers v5 is FINALLY out!!

voice ai is about to have a moment i don't think people get how rapidly the experience changes as latency approaches 0

Introducing our latest open model: MedASR 🔬Speech to text model 🏥for healthcare-based voice applications 🤗available in Hugging Face ⚡️run with transformers Download right now huggingface.co/google/medasr

@kadirnardev ai.meta.com/research/publi…

@eustachelb Paper?

Boom! Big release by Meta in the audio game: Sam Audio and Perception Encoder Audiovisual. "The core innovation in SAM Audio is the Perception Encoder Audiovisual engine"... and it's supported day 0 in transformers!

Check out the models huggingface.co/collections/fa…

Check out the release blog post! huggingface.co/blog/transform…

Without it, using a new model often means diving into a complex research codebase, deciphering experimental features, and dealing with huge engineering overhead. Most people simply give up, and end up paying for a closed solution. transformers solves this. Learn one paradigm. Trust the integrations. Build with any open model. Move faster, not slower. It is the source of truth for open models, and I’m incredibly proud to help build it. Here’s to open source, Hugging Face, and the entire community. ❤️‍🔥

transformers v5 is out! 3 mill daily downloads, one of the core foundations of the open AI ecosystem. 4y after v4, this milestone reminds us why we build OS AI. It’s genuinely fun to work on, but as technicians, we need to think about the impact of the tools we’re creating.