Kwanghee Choi (@juice500ml) profile picture

PhD student, working on speech AI with David Harwath @utsaltlab (@UTAustin) and David R. Mortensen @dmort27 (@LTIatCMU).

Joined August 2023
Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)Photo by Kwanghee Choi (@juice500ml)
Kwanghee Choi retweeted
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
261
826
288
7,382
1,046,886
Kwanghee Choi retweeted
๐Ÿ“ข New paper: RUBATO: Decoupled Clocks Allow Full-Duplex Models to Think as Fast as They Can Current full-duplex models tie thinking to the conversation's pace. RUBATO decouples them, giving each stream its own clock, so thinking runs as fast as the hardware allows.
3
21
4
114
5,591
Kwanghee Choi retweeted
(1/7)๐Ÿ“ข New paper! Speech AI is getting better at hearing and speaking many languages. But what if the input mixes two languages, or is in one the model never learned? It guesses one language it knows and imposes that language's sounds on everything. We find why, and fix it ๐Ÿงต
2
9
1
28
855
๐Ÿค– How well do agents learn from experience? We study agents that amortize experience into reusable artifacts and introduce Agent Plasticity to measure how efficiently they improve on held-out environments. Top performer โ‰  most efficient learner.
17
55
16
273
37,235
Kwanghee Choi retweeted
Excited to share our new paper to be presented at SLT 2026 in Palermo! ๐Ÿ‡ฎ๐Ÿ‡น ๐ŸŽ‰ We show that phone segmentation and recognition can be solved jointly by manipulating SSL representations using less than a minute of labeled speech. (1/n)
1
10
25
1,763
3 papers submitted & accepted at #SLT2026 ๐ŸŽ‰ Great to see that phonological vector arithmetic is indeed effective across several new applications. Bunch of computational linguistics this time, see you in Palermo!
1
2
16
712
Foundation models are built in various stages, from pre-training to post-training and beyond. We still lack a principled understanding of how these stages interact. Excited to co-organize โ€œTransitioning from Pre-Training to Post-Trainingโ€ at NeurIPS 2026 in Sydney. Submit your work by Aug 29!
announcing the NeurIPS 2026 workshop on *Transitioning from Pre-training to Post-training*!๐Ÿ“ come bring your theory, empirics, & methods to enable better foundations, mechanics, & evaluation for LLM training! submit your work by 29th August!๐Ÿ—“๏ธ pretrain2posttrain.github.ioโ€ฆ ๐Ÿงต
2
9
2
31
4,483
See you at San Diego on Sunday 14:00-15:30, Poster Session B! I'll be presenting the [b] = [d] - [t] + [p] paper ๐Ÿค“
4 papers submitted & accepted at ACL 2026! ๐ŸŽ‰ So grateful to work alongside & learn from amazing minds, pushing the boundaries of speech technologies, machine learning, and computational linguistics. See you in San Diego!
4
271
3 papers accepted at Interspeech 2026! ๐ŸŽ‰ Proud to see my role slowly transitioning from a mentee to a mentor. Huge thanks to my collaborators for letting me be a part of their journey!
3
13
868
Kwanghee Choi retweeted
What if you had nano-banana for audio? AudioChat is a multi-modal LM that performs fine-grained understanding, generation, and editing of multi-source scenes By diffusing continuous latents, it generates 48khz stereo edits with great input adherence: wanchichen.github.io/audiochโ€ฆ
6
24
3
168
18,466
Kwanghee Choi retweeted
โœจNew paperโœจ We find script (e.g. Cyrillic, Latin) to be a linear direction in the activation space of Whisper, enabling transliteration at test-time by adding such script directions to the activations โ€” producing e.g. Cyrillic Japanese transcriptions.
1
1
3
265
4 papers submitted & accepted at ACL 2026! ๐ŸŽ‰ So grateful to work alongside & learn from amazing minds, pushing the boundaries of speech technologies, machine learning, and computational linguistics. See you in San Diego!
9
1
43
2,290
Thanks a lot for the interest in our work! Here's the recording for people who missed the seminar: youtu.be/DtFYKvNo9IQ
๐’๐ž๐ฅ๐Ÿ-๐ฌ๐ฎ๐ฉ๐ž๐ซ๐ฏ๐ข๐ฌ๐ž๐ ๐’๐ฉ๐ž๐ž๐œ๐ก ๐Œ๐จ๐๐ž๐ฅ๐ฌ ๐š๐ซ๐ž ๐๐ก๐จ๐ง๐จ๐ฅ๐จ๐ ๐ข๐œ๐š๐ฅ ๐•๐ž๐œ๐ญ๐จ๐ซ ๐Œ๐š๐œ๐ก๐ข๐ง๐ž๐ฌ! ๐Ÿ—ฃ๏ธ Excited to be giving an invited talk this Thursday (March 19th, 3pm Amsterdam time)! Huge thanks to @mariannedhk at Univ. of Amsterdam for the invite ๐Ÿ™
7
375
๐’๐ž๐ฅ๐Ÿ-๐ฌ๐ฎ๐ฉ๐ž๐ซ๐ฏ๐ข๐ฌ๐ž๐ ๐’๐ฉ๐ž๐ž๐œ๐ก ๐Œ๐จ๐๐ž๐ฅ๐ฌ ๐š๐ซ๐ž ๐๐ก๐จ๐ง๐จ๐ฅ๐จ๐ ๐ข๐œ๐š๐ฅ ๐•๐ž๐œ๐ญ๐จ๐ซ ๐Œ๐š๐œ๐ก๐ข๐ง๐ž๐ฌ! ๐Ÿ—ฃ๏ธ Excited to be giving an invited talk this Thursday (March 19th, 3pm Amsterdam time)! Huge thanks to @mariannedhk at Univ. of Amsterdam for the invite ๐Ÿ™
1
5
3
17
2,087
๐Ÿงต Together, both papers take a step beyond the usual "what info do S3Ms encode" probing paradigm. We aim to answer how is that info actually encoded geometrically? Come see for yourself Thursday! ๐Ÿ‘€ Slides: docs.google.com/presentationโ€ฆ
1
1
119
Huge thanks for my wonderful coauthors, Eunjung and Cheol-jun, and my two favorite Davids, Mortensen ๐Ÿ‘ and Harwath ๐Ÿค  โ€” best advisors I could ask for ๐Ÿ™ Can't wait to see what we cook up next! ๐Ÿš€
105