SeungHeon Doh Profile
SeungHeon Doh

@SeungHeon_Doh

Followers
777
Following
1K
Media
52
Statuses
228

LLM + Music | Postdoctoral researcher @ KAIST | Intern @SonyAI_global | Previously an intern @Adobe, @BytedanceTalk, @Naver, @Chartmetric.

Joined June 2022
Don't wanna be here? Send us removal request.
@SeungHeon_Doh
SeungHeon Doh
30 days
Can a single framework (semantic IDs or embedding similarity) fully satisfy user needs in conversational music recommendation? As a solution, we present TalkPlay-Tools. - Demo: https://t.co/4EBs8bhDBp - Paper: https://t.co/GXiYlLk5bS - Code: https://t.co/vFW0CnBRsi
1
4
22
@keunwoochoi
Keunwoo Choi
12 days
NLP4MUSA 2026 officially announced a new challenge on conversational music recommendation. this year's challenge is based on TalkPlayData-2! i'm so honored to have contributed to this timely challenge.
1
2
11
@SeungHeon_Doh
SeungHeon Doh
11 days
@keunwoochoi
Keunwoo Choi
11 days
tomorrow (10/23 Thu), 4:30PM EST. my talk on TalkPlay representing https://t.co/CTRcCcwzvW. at RIT (Rochester) / on Zoom https://t.co/9vr6coi9Lw
0
0
2
@SeungHeon_Doh
SeungHeon Doh
13 days
This year: 1) NLP4MusA is expanding ISMIR to EACL 2) the Conversational Music Recommendation Challenge has been newly established! - Challenge Dataset: https://t.co/5eYpbEDKeJ - Baseline Code: https://t.co/Yv1ji0FERn - Evaluation Code:
0
1
3
@SeungHeon_Doh
SeungHeon Doh
13 days
Call for Papers: NLP4MusA 2026 We welcome work at the intersection of language, music, and audio — from music understanding and recommendation to creative generation. Submission deadline: Dec 19, 2025 Details & submission:
Tweet card summary image
sites.google.com
🎵 First Call for Papers: 4th Workshop on NLP for Music and Audio (NLP4MusA 2026) Co-located with EACL 2026, Rabat, Morocco & Online | March 24–29, 2026 Shared Task: Conversational Music Recommenda...
1
3
7
@affige_yang
Yi-Hsuan Yang
21 days
New paper on symbolic music generation: first envisioning a theme, then composing the segments one-by-one in a flexible order. demo - https://t.co/YC0YuHcwMc code -
Tweet card summary image
github.com
NeurIPS 2025 AI4Music Workshop: Segment-Factorized Full-Song Generation on Symbolic Piano Music - eri24816/segmented-full-song-gen
@ArxivSound
arXiv Sound
25 days
Ping-Yi Chen, Chih-Pin Tan, Yi-Hsuan Yang, "Segment-Factorized Full-Song Generation on Symbolic Piano Music,"
0
4
17
@_reachsumit
Sumit
1 month
TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling @SeungHeon_Doh et al. position an LLM as an end-to-end recommendation system that interprets user intent and plans tool invocations for music discovery. 📝 https://t.co/fjp8PYMFUF 👨🏽‍💻 https://t.co/rCdjfes0yt
Tweet card summary image
arxiv.org
While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited,...
1
1
7
@SeungHeon_Doh
SeungHeon Doh
30 days
This work was carried out together with co-first author @keunwoochoi and @juhan_nam. We plan to present it at the NeurIPS AI4Music Workshop!
0
0
0
@SeungHeon_Doh
SeungHeon Doh
30 days
Through a simple conversational music recommendation experiment, we confirm that the proposed tool calling framework outperforms zero-shot LLM recommendation models. (When both use Chain-of-thought)
1
0
0
@SeungHeon_Doh
SeungHeon Doh
30 days
Tool calling doesn’t rely on a single search model; instead, it uses multiple search models sequentially, supporting a retrieval–reranking framework.
1
0
0
@SeungHeon_Doh
SeungHeon Doh
30 days
We started this project with the idea that different user queries require different models: numerical values like tempo or popularity suit SQL, artist or album searches work best with BM25, personalization fits CF models, and timbre-based searches rely on audio embeddings.
1
0
0
@SeungHeon_Doh
SeungHeon Doh
30 days
TalkPlay-Tools consists of Music Recommendation Agents and an external environment. We use various music search and personalization models as tools.
1
0
0
@zacknovack
Zachary Novack
2 months
I am officially on the job market!! Looking for industry and faculty/postdoc starting Fall 2026, especially in Audio/Music Gen and Audio-LLMs 🥳 I'll be at #ISMIR2025 this week (super pumped!), where I'll present 3 papers in AI Music eval and datasets if you want to chat! 🇰🇷
8
11
83
@SeungHeon_Doh
SeungHeon Doh
2 months
A New Conversational Music Recommendation Dataset with (1) User Profile, (2) Conversation Goal, (3) Chain of Thought, and (4) Multimodal Item Representation. https://t.co/iONg8vZUEH
@keunwoochoi
Keunwoo Choi
2 months
TalkPlayData2 paper is on arXiv now! https://t.co/yVpFCpKxai
0
2
16
@DBralios
Dimitrios Bralios
2 months
Great audio AEs/codecs exist, but when you need structured latents or a tweaked bottleneck for a downstream task (e.g. generation), retraining is expensive & brittle. We Re-Bottleneck👇
@ArxivSound
arXiv Sound
2 months
Dimitrios Bralios, Jonah Casebeer, Paris Smaragdis, "Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders,"
1
5
20
@SeungHeon_Doh
SeungHeon Doh
2 months
New to ISMIR? Join the Newcomer Squad! Have you participated in ISMIR multiple times? Become a Newcomer Leader!
@ISMIRConf
ISMIR Conference
2 months
Join the ISMIR2025 Newcomer Squad – Connect, Learn & Enjoy! 📷 Is this your first ISMIR and you’re not sure how to make the most of it? Or are you a seasoned attendee who’d love to give back and welcome newcomers? Either way, the Newcomer Squad at ISMIR2025 (South Korea!) is
0
1
8
@jesseengel
Jesse Engel
3 months
Realtime interactive generative models FTW! Announcing a new 🌊 of details and features for Magenta RealTime, the open weights live music AI model from GDM! * Live Jamming with audio input 🎤🎸🎵 * Personalize your own models 🔧 * Tech report 📜 Links below in the 🧵...
81
170
1K
@zacknovack
Zachary Novack
3 months
Suno + Veo 3 generate highly similar versions of popular songs purely based on *phonetically* similar gibberish lyrics?!?! Presenting Bob’s Confetti: Phonetic Memorization Attacks in Music and Video Generation 🔊: https://t.co/Rztf9TNhI7 📖: https://t.co/H7BgnUkCvD 🧵1/n
2
25
90