Hang Le

PhD candidate in Speech and Language Processing

[laurent besacier @laurent_besacie]

I refreshed my ASR course (<100 slides): fewer HMMs, more speech tokens, speech LLMs and full-duplex models. Claude helped with the figures and more. Feedback welcome! https://t.co/BrzLq3Y8Q2

[laurent besacier @laurent_besacie]

Applications to ALPS 2027 are open !!!

[AI at Meta @AIatMeta]

SeamlessStreaming is an AI translation model that can deliver state-of-the-art results on streaming translation with <2 seconds of latency. One core piece of our latest Seamless Communication research work by teams at FAIR.

More on this project ➡️ https://t.co/pPgkrTvryg

[laurent besacier @laurent_besacie]

Opt for lang-specific encoders for real-world applications in lang like French. LeBenchmark 2.0 is here with 3 new pre-trained speech SSL models (until 1B), tested on 6 tasks, powered by Jean Zay #HPC from @Genci_fr.
Paper: https://t.co/sqA1nMAnMz
Models: https://t.co/NWB4ZcmuYx

[Juan Pino @juanmiguelpino]

Congratulations to the team on this exciting release featuring an all-in-one massively multilingual multimodal machine translation model! What I'm particularly excited about for this work👇

[didier_schwab @didier_schwab]

Projet ANR Pantagruel accepté :) Construction et évaluation de giga modèles de langue multimodaux et inclusifs (écrit, oral, pictogrammes) pour le français général et clinique https://t.co/K63U17ld3f

[laurent besacier @laurent_besacie]

Presented this week at ACL (from NLP@NLE): Memory-efficient NLLB-200: language-specific expert pruning of a massively multilingual machine translation model - Naver Labs Europe https://t.co/N2UezdJAf5

[laurent besacier @laurent_besacie]

Presented this week at IWSLT@ACL: NAVER LABS Europe’s Multilingual Speech Translation Systems for the IWSLT2023 Low-Resource Track (were ranked 1st on the 2 lang pairs they participated in this track) - https://t.co/gRf9NhrtM1 with @mzboito I.calapodescu, a.berard & e.gow-smith !

[Matthias Gallé @mgalle]

We are doing🏔️ALPS 🏔️again, 4⃣th edition

A week full of insightful lectures, knowing fellow researchers and dreaming of joint projects

≠ this time: we actually meet in the Alps. Close to a French National Park, at 1500m, surrounded by mountains

https://t.co/pNqBUzvo6g

[Hang Le @formiel]

@JohnTsiamas @geiongallego @costajussamarta Thanks Ioannis for the kind words! Your work is also super interesting (have been following closely your team’s publications for a while :D @mtupc1).

[Hang Le @formiel]

A nice combination of strong techniques for speech translation, including an interesting extension of our Siamese pre-training!

[Ioannis (Yiannis) Tsiamas @JohnTsiamas]

Excited to share our submission to #IWSLT 2023. We utilized large speech & text foundation models, and pretrained them with Optimal Transport to adapt the speech representations to the text space, before supervised ST fine-tuning.

Check our method here: https://t.co/8iMJror1vU

[MT Group at FBK @fbk_mt]

Our Pick of the week: Phuong-Hang Le et al., "Pre-training for Speech Translation: CTC Meets Optimal Transport" by @mgaido91
https://t.co/AnbGsxty2k
#NLProc #optimaltransport #CTC #speechtranslation

[Hang Le @formiel]

@mgaido91 Thanks Marco for your interest in our work!! That's indeed interesting to experiment. I may give it a try during the summer.

[Marco Gaido @mgaido91]

Many papers recently exploit optimal transport-based losses, and it is great to see how this can be done to improve ST models in https://t.co/wCMLPIGOZS. I would have been curious to see results on the Conformer model and with CTC compression!

[Hang Le @formiel]

@lorelupo Thanks a lot, Lorenzo!!

[Hang Le @formiel]

Happy to share our recent work on speech translation (ST): https://t.co/xS5hsAorRS, which will be presented at ICML 2023. We make two contributions: 1) showing that CTC can reduce the modality gap in ST pre-training... 1/2

[Hang Le @formiel]

@khainb_ml Many thanks Khải @khainb_ml! Congrats on your paper as well!! Unfortunately I won't be able to attend due to visa issues :( Enjoy Hawaii and the conference. Hope to meet you at another venue in the future.

[Hang Le @formiel]

@_akhaliq Thanks @_akhaliq for sharing our work! The code and pre-trained models are currently available at https://t.co/xTVNYCokfN

[AK @_akhaliq]

Pre-training for Speech Translation: CTC Meets Optimal Transport

abs: https://t.co/xirkpYfEgR

[Hang Le @formiel]

and 2) proposing a pre-training method using optimal transport for aligning speech and text sequences. Joint work with @AnnFirst111 @ChanghanWang @juanmiguelpino Benjamin Lecouteux, @didier_schwab.