Our Embedded AI group is looking for its next Research Director. We're researching and commercialising at the intersection of Embedded and distributed deep learning, GenAI and speech technologies.

Headed to @interspeech.bsky.social next week to present the Samsung AI Cambridge Speech Team work—if you see someone in a SpeechBrain or Samsung AI hoodie, that’s me. Come say hello!

📢 Join our Conversational AI Reading Group! 📅 Thursday, April 17th | 11 AM - 12 PM EST 🎙 Speaker: Titouan Parcollet (@tparcollet.bsky.social) from Samsung AI Center Cambridge 📖 Topic: "Unsupervised on-device adaptation of a speech recogniser and the Pitfalls of "SpeechLLM" evaluation"

Your streaming speech recognizer is probably mathematically flawed, degrading its accuracy. Ask me to explain how to fix this next week in the Thursday morning poster session at #ICASSP, or look at ieeexplore.ieee.org/abstract/doc...

Most PhD in CS are not actually PhD. We should really start splitting things properly with those seeking excellence in engineering and those seeking excellence in science. Less useless publications, clearer path towards hiring.

As a French, I do not understand how the US aren’t in a revolution. Macron can only dream of the situation, as we would burn the country for 1/5 of what Musk is doing right now. If Musk and Trump get away with this, it will be an insane signal for our EU governments. Don’t let this happen folks.

Repeat after me: “AI will not save us from the major moral and environmental issues that we are facing”.

110B invested in France for AI. I really wish that we stop burning our cash like that on AI and, instead, focus on truly important issues like health or energy…

Congratulations to NatureLM-audio authors David Robinson, Marius Miron, @mhagiwara.bsky.social, and Olivier Pietquin on this huge accomplishment. We're looking forward to presenting at #ICLR2025 in April.

Walking around Cambridge …

They've been burning the world done for many generations already. This new gen of "bros" known as "Tech bros" are just more visible than our usual oil/mining/pharma/agri bros. And by the way, the problem does not come from "bros" but from the power you can get from having absurd amounts of money.

@ovhcloudfr.bsky.social vous n’auriez pas une instance de stockage optimisée à prêter au projet open source @speechbrain.bsky.social pour la préparation / release d’un dataset sur @hf.co de reconnaissance de la parole de 25.000 heures?

It looks like there isn’t a single “rich” country where the politics and the economy aren’t moving forward in favor of the rich and against the people with more and more violence. It seriously looks like the agony of an old system, and privileged people just don’t want to accept it.

The silence around Long Covid is deafening because speaking up requires - Fighting waves of misinformation to understand your condition - Accepting an incurable illness as your new reality - Risking relationships, income & facing stigma, all for little hope of support. Most simply suffer quietly.

I just deleted my Facebook account.

Please conferences… don’t ban papers from being published on Arxiv before submission… as long as arxiv exists, this is a terrible idea!

The worst thing that can happen to a research field is to attract too many investors and too much media coverage - rip AI.

Pretty sure that it is no better in the industry.

This. We must fund research on long-covid / Chronic Fatigue Syndrome. People with these diseases are properly invisible, because they cannot be physically visible anywhere.

Transparency isn’t sufficient in most cases. In the EU, legal concepts to force companies to not harm people exist. Yet, it does not prevent them from doing it. What we need is a proper legal framework describing disproportionate energy consumption as a crime. Then we will be able to act.

Today, we’re introducing NatureLM-audio: the first large audio-language model tailored for understanding animal sounds. arxiv.org/abs/2411.07186 🧵👇

I just discovered Earth Species Project. That’s a freakin good way of applying AI for the greater good. No AGI bullshit, no social harm, just knowledge and collaboration with nature.

He smoll, but he beautifol.

Yes. 100%. If you work in a place not providing you at least half of the points in that list, and if you are able to, change, and quickly, you only have one life.

People who say that the private sector is much more efficient money-wise than the public sector have either never worked in a medium—to large company or startup or have never been employed full-time as a public servant.

Massive work. Now my take: if you need to do such a thing for a conference that ends up being a random process anyway, it means that it is time to re-think it entirely. NeurIPS, ICLR and the like are an excellent peak at how terrible this field has become. Mass production of mostly shit science.

“Only the richest 10 per cent of households in the country can afford to buy an average-priced home in England” — capitalism gone right. www.independent.co.uk/news/uk/home...

About to drop a 25,000 hours dataset for English ASR with allowed commercial use and both clean / noisy and read / spontaneous speech. It will also contain a 2.5k hours subset and validation / test sets. Hopefully on @speechbrain.bsky.social :)

“I am right, the parliament is wrong, and you are wrong, you just don’t understand” — basically what Emperor Macron just said to the French people during an official TV communication. The dude has already forgotten the performance of Gojira during the Olympic Ceremony it seems.

We are excited to announce the launch of ML SUPERB 2.0 (multilingual.superbbenchmark.org) as part of the Interspeech 2024 official challenge! We hope this upgraded version of ML SUPERB advances universal access to speech processing worldwide. Please join it! #Interspeech2025

To anyone using the YODAS dataset for ASR (huggingface.co/datasets/esp...) be aware that many samples with a given language flag may actually be of another one! (e.g. French instead of English!)

French universities are facing a risk of general collapse or serious service degradation due to years and years of right-minded policies. Do French people care? Absolutely not, they don’t even know what universities are. They can’t even comprehend how lucky they are to get free education.

🎓Academia or the industry 💸? I wrote a detailed point of view on Twitter a few months ago, so maybe I should share it here again. I think that most things are still true, the only slight change would be linked to the GenAI bubble, but only time will tell. www.darnault-parcollet.fr/documents/Ba...

Thanks to the help from the @hf.co people, we fixed the hub of the People's Speech ASR dataset (28k hours of speech). For those wondering, the "clean" set are all samples with a CER lower than 20% between the transcription and the output of a pre-trained ASR model. (This explains the terrible WER)