Thibault Bañeras-Roux

PhD - NLP & Speech Proc. Currently at @UCLouvain, @CENTAL_ucl

GitHub:

[Thibault Bañeras-Roux @BanerasRoux]

Voici HATS !🎩

Grâce aux 143 personnes qui ont participé à mon expérience, j'ai pu obtenir 7150 annotations et construire le data set HATS que nous diffusons librement afin que la communauté puisse évaluer les métriques de reconnaissance de la Parole🎙️

[Richard Dufour @r_dufour]

⚠️ [Appel à Participation]
Campagne évaluation DEFT 2024 ⚠️

📝Tâche : Réponse automatique à des QCM issus d'annales d'examens de pharmacie

🌐 Plus d'infos : https://t.co/bAvXTfe4oF

🚀N'hésitez pas à participer !

[Thibault Bañeras-Roux @BanerasRoux]

@ParcolletT I have the feeling that most of them don't even listen to scientifics.

[Thibault Bañeras-Roux @BanerasRoux]

@ParcolletT We expect politicians to be expert in everything but it is not possible. This is one of the main problem of representative democracy.

[Laboratoire Parole et Langage @LPL_lab_Aix]

📣🎉 Lancement officiel du Master trilangue #MASTRI #créole #francais #anglais à @UniofSeychelles en collaboration avec @univamu et avec le soutien de l'Ambassade de France aux Seychelles. Projet #FSPI-R coordonné par Sibylle Kriegel du @LPL_lab_Aix.
➡️ https://t.co/ByHvrfRpiE 👏

[Richard Dufour @r_dufour]

🚀 🏥 Very proud to announce BioMistral, a collection of open-source pre-trained LLMs for the medical domain

📰Arxiv: https://t.co/Zarwhb807S
🏥 BioMistral 7B model: https://t.co/h00IOmXC66

More info: https://t.co/eiFj4bk8b5

@CNRSinformatics @LaboLS2N @taln_ls2n @LabrakYanis

[Thibault Bañeras-Roux @BanerasRoux]

Merci à tous les participants qui ont participé à cette expérience qui n'était pas simple !

[Thibault Bañeras-Roux @BanerasRoux]

Voici HATS !🎩

Grâce aux 143 personnes qui ont participé à mon expérience, j'ai pu obtenir 7150 annotations et construire le data set HATS que nous diffusons librement afin que la communauté puisse évaluer les métriques de reconnaissance de la Parole🎙️

[Thibault Bañeras-Roux @BanerasRoux]

Lien du GitHub pour obtenir le data set et l'outil pour évaluer automatiquement votre métrique : https://t.co/6BXeHVpo9G

Papier : HATS: An Open data set Integrating Human Perception Applied to the Evaluation of Automatic Speech Recognition Metrics

[Thibault Bañeras-Roux @BanerasRoux]

From a semantic aspect, one could also use semantic metrics with language-model embeddings.

We recently brought evidences that SemDist or BERTScore shows a better correlation with human perception for French.

https://t.co/NWgXyqQTfF

[AlphaCephei @alphacep]

Not just WER is important, nice idea here. Similar to this, NVIDIA adds BLEU score to ASR models.

https://t.co/BBONQjyVBi

[Josh Meyer @_josh_meyer_]

Coqui is shutting down.

It's sad news to start the new year, but I want to take a minute to recognize everything we accomplished and thank the great people who made it possible.

First things first: the Team

I'm honored to have worked with such brilliant, dedicated, and inspiring individuals. We were a small team, but we left our scratch on the earth's crust. Our accomplishments stand on their own, but when you remember we were just a rag-tag team with limited compute... now that's special.

Big tech had orders of magnitude more compute, data, and researchers, but we gave them a run for their money. We didn't just replicate the state-of-the-art... we created it! That wouldn't have been possible without this exact team.

We were spread across five continents, native languages, and backgrounds... and we built something great. I'm sure that we built great tech because of that mix of perspectives.

I will deeply miss our team, but I'm also excited to see what they do next. Whoever gets them on-board will be a lucky duck :)

What we accomplished

Way back in 2016, it all began as the Machine Learning Group at Mozilla. First was DeepSpeech, then Common Voice and TTS. Crazy how far the field has come since then. We spun out as Coqui in 2021 in order to add rocket fuel to our mission.

One of our biggest accomplishments at Coqui was XTTS. The state-of-the-art took a huge leap forward when we openly released model weights for XTTS v1... and v2 was even better! I'm thrilled to see where AI is heading, and proud that we could make some of that progress available to everyone.

Here's a tiny snapshot of what we accomplished at Coqui:

✅ 2021: Coqui STT v1.0 release. Coqui Model Zoo goes live. SC-GlowTTS released.
✅ 2022: YourTTS goes viral. Tons of open-source releases. Building the team.
✅ 2023: Coqui Studio webapp and API go live. First customers. XTTS open release.

I can confidently say that we pushed the state-of-the-art for generative speech technology... before it was called "generative" :)

Thank you

It took a village to make Coqui possible, and I want to thank everyone who gave us a shot.

The real rockstars are the team, as I said above. Thank you!

A huge thanks to the community. You have always been our core. From the Mozilla days on IRC to the current Discord server. The community has contributed, supported, and made building in the open a joy. Thank you all!

Thank you to our investors. Coqui simply wouldn't have been possible without you. You believed in us before anyone else; you took a chance on us. More than just an investment, your thoughtful insights and discussions made Coqui a better company and a better product. I'm extremely grateful for your support. Thank you!

Thank you to our customers. Everything we built was for you, and I hope we managed to give you something you loved. Especially thank you for your feedback: both the good and the bad. We did our best to hear you and build you something better everyday. Thank you!

Lastly, thank you to our partners over the years. It's a long list of great folks I've been lucky enough to collaborate with. We worked on open science, open code, and open models. From joint research to hackathons, it was a blast! To the great folks at HuggingFace, Mozilla, Masakhane, Harvard, Indiana University, Google, MLCommons, Landing AI, NVIDIA, Intel, and Makerere University... thank you! Forgive me if I've left anyone out.

What's next

I can't yet say what comes next... but generative AI in 2024 is going to be bigger than ever. Generative voice will only get better, faster, cheaper, and easier to fine-tune... open-source will be a huge part of that.

Speaking of open-source... Coqui TTS is on Github. Do something awesome with it!

Thank you all 💚

https://t.co/yXcQY0g9wa

[Laboratoire Parole et Langage @LPL_lab_Aix]

📣📣📣 Appel à participant.e.s pour une expérience de perception de sons au @LPL_lab_Aix à Aix-en-Provence !
Contact : [email protected]

[Gael Varoquaux 🦋 @GaelVaroquaux]

Avec la #LoiImmigration, le gouvernement manie la xénophobie, et veut inscrire la discrimination dans la loi.

C'est le programme de l'extrême droite, un programme de division et non de construction, un programme qui met notre démocratie sur une pente dangereuse.

[Lê Nguyên Hoang bsky.app/profile/science4all.org @le_science4all]

Le plus long mot du dictionnaire français résume l'actualité politique française. #Anticonstitutionnellement
https://t.co/Eks5encXQx

[Thibault Bañeras-Roux @BanerasRoux]

"Un ordinateur créé avec du tissu cérébral humain excelle en reconnaissance vocale"

[Trust My Science @TrustMyScience]

Brainoware, fusionnant tissu cérébral humain et électronique, promet des avancées en IA et en neurosciences tout en soulevant d'importantes questions éthiques. https://t.co/kH91en8wkZ

[Thibault Bañeras-Roux @BanerasRoux]

@charlylamothe du LIS/INT nous a présenté ses travaux très intéressants au Laboratoire d'Informatique d'Avignon !

Sa thèse a porté sur l'analyse cérébrale, et ses liens avec le traitement naturel et automatique de la parole.

Merci à lui 🙂

[CNRS Sciences informatiques @CNRSinformatics]

#SemaineTAL #RésultatScientifique 🔎 | Le projet MALADES coordonné par @r_dufour et Pantagruel, coordonné par @didier_schwab visent à répondre aux enjeux scientifiques et sociétaux soulevés par l’émergence des grands modèles de langue
➡https://t.co/grOJNeSZQ9
🤝@LaboLS2N @LIGLab

[New Scientist @newscientist]

A computer science student has deciphered a word on a badly charred and tightly rolled papyrus scroll unearthed in the Roman town of Herculaneum, which was buried in a volcanic disaster in AD 79. https://t.co/oGwvuHExVy

[Adel @adelmoumen_]

👋 Hey Twitter!

I'm looking for a PhD position in efficient deep learning, with a special interest in speech recognition.

I've actively contributed to @SpeechBrain1 and want to continue exploring open-source initiatives.

CV: https://t.co/UJQvmqpgCu

DMs open!

Please RT! 🙏🏼

[Thibault Bañeras-Roux @BanerasRoux]

@ArthurAmalvy @r_dufour @VincentLabatut Félicitations !