Andy Ehrenberg

Research @MistralAI

[Mistral AI @MistralAI]

🔊Introducing Voxtral TTS: our new frontier open-weight model for natural, expressive, and ultra-fast text-to-speech

🎭Realistic, emotionally expressive speech.
🌍Supports 9 languages and accurately captures diverse dialects.
⚡Very low latency for time-to-first-audio.
🔄Easily adaptable to new voices

[Andy Ehrenberg @andy_ehrenberg]

@ricburton @ichati_uchiha @AaronSchnacky drake's estero

[Mistral AI @MistralAI]

We’ve raised €1.7B to accelerate technological progress with AI!

This Series C funding round, led by @ASMLcompany, fuels Mistral AI scientific research to keep pushing the frontier of AI to tackle the most critical technological challenges faced by strategic industries.

[Soham @sohamg121]

The Voxtral tech-report is up!
https://t.co/1qzV3ShvDE

We release these models with a permissive Apache 2.0 license. Feedback is welcome! We have a lot more cooking, this is just the beginning.

[Andy Ehrenberg @andy_ehrenberg]

More coming soon!

[Mistral AI @MistralAI]

Introducing the world's best (and open) speech recognition models!

[Mistral AI @MistralAI]

Introducing Mistral Small 3.1.
Multimodal, Apache 2.0, outperforms Gemma 3 and GPT 4o-mini.
https://t.co/BHLAAaKZ9w

[Andy Ehrenberg @andy_ehrenberg]

@hingeloss I don’t have experience with jetstream-pytorch but have had some headaches with torch on XLA in the past…

[Andy Ehrenberg @andy_ehrenberg]

@alex_conneau Question about the ASR evals - does 16-shot there just mean 16 audio-transcription examples in a prompt to show formatting/etc? Are the 16 examples fixed or do they vary depending on the audio's language/setting?

[Andy Ehrenberg @andy_ehrenberg]

Great team effort that required rethinking everything from data storage to hardware to training strategies. Happy with the result and really excited about the future projects unlocked by the changes to our infra.

[Robert McHardy @robert_mchardy]

Very excited to share our pre-print on how we (@AssemblyAI) built our new multilingual speech-to-text model, Universal-1, by massively scaling unsupervised and supervised data to many millions of hours of audio.

Check out the pre-print for more details:
https://t.co/zzXMgtsBwQ

[AssemblyAI @AssemblyAI]

Introducing Universal-1, our most powerful speech recognition model to date.

Trained on over 12.5 million hours of multilingual audio data, Universal-1 achieves best-in-class speech-to-text accuracy across English, Spanish, French, and German.

https://t.co/xgrE6EhmJu

[AssemblyAI @AssemblyAI]

We’re excited to announce that AssemblyAI has raised $50M in Series C funding to help further our vision of building superhuman Speech AI models that unlock the power of voice data for every organization.

The round, led by @Accel with participation from @KeithBlock and Smith Point Capital, @insightpartners, @danielgross and @natfriedman, and @ycombinator brings our total funding to $115M.

https://t.co/IrIMrmd20H

[Robert McHardy @robert_mchardy]

Very cool to see Google Cloud covering how we're using TPUv5e and Jax at @AssemblyAI

https://t.co/ZZK37cnoEq

[Andy Ehrenberg @andy_ehrenberg]

Another great example of the power of named axes, this time with a very nice guide on doing FSDP with JAX

[Enrique Piqueras @epiqueras1]

Behold, the power of JAX and XLA!

https://t.co/XrFFC2TvAp

[Andy Ehrenberg @andy_ehrenberg]

@ednewtonrex Time to start a jam band

[Andy Ehrenberg @andy_ehrenberg]

@ednewtonrex Maybe you have some insight here

[Andy Ehrenberg @andy_ehrenberg]

Who’s working on p(stem|other stems for a song, [maybe the target instrument])? Lot of musicians would find this more useful than p(all tracks|text description) if there’s sufficient data out there

[Andy Ehrenberg @andy_ehrenberg]

Fun making this contribution! Has made it very fast to eval our models against Whisper.

[Sanchit Gandhi @sanchitgandhi99]

Thanks to @andy_ehrenberg, 🤗 Transformers now has support for JAX/Flax Whisper 🚀

With minimal code changes, you can run Whisper up to **10x faster** in a Google Colab

Check out the notebook: https://t.co/aSIq8UFRDW

[Andy Ehrenberg @andy_ehrenberg]

@spolu Are you saying just doing SL on trajectories from a bigger model from when the bigger model underwent rlhf, or are you using rl to update the smaller model?

[Andy Ehrenberg @andy_ehrenberg]

@spolu I say this cause you’re talking about distillation but then saying that ChatGPT gained skills over what it would have been distilled from in this hypothetical setup