Nithya Nadig Shikarpur

Excited to announce our ICASSP 2026 paper "Stemphonic: All-at-once Flexible Multi-stem Music Generation" ! w/ @__gzhu__, @j_p_caceres, @huangcza, and @NicholasJBryan 🔊Demo stemphonic-demo.vercel.app 📰Paper arxiv.org/abs/2602.09891 More details in🧵

Thrilled to announce “MIDI-LLM: Adapting LLMs for Text-to-MIDI Music Generation” w/ @huangcza and Yoon Kim! 🎸 Live Demo midi-llm-demo.vercel.app 💻 github.com/slSeanWU/MIDI-… 🤗 huggingface.co/slseanwu/MIDI-… From a text prompt, it generates MIDIs you can edit directly in a DAW 🧵

How do we design real-time audio jamming models, and do we have the right data? Our new arXiv preprint studies the design and its tradeoffs. Paper: arxiv.org/abs/2510.22105 Audio samples & code: lukewys.github.io/stream-music-g…

This is happening tomorrow from 9am to 12 pm and 1 pm-4 pm in East Ballroom C! You can come and interact with our model too!! 🎤 Hope to see you there :) #NeurIPS2024

Work done with @huangcza at @Mila_Quebec and Université de Montréal!

While the model presents exciting directions for creative explorations and human-AI partnerships, this work notes the experiences of musicians to inform the model's future development! More information on the generative model: x.com/NithyaIsMe/sta…

We studied some musicians interacting with GaMaDHaNi, a generative model for Hindustani vocal music 🎤! I am so excited to present this work at @NeurIPSConf @ML4CDworkshop this week! 📝arxiv.org/abs/2411.13846 🎧cedar-decade-974.notion.site/Example-Videos… An example interaction with the model:

I'm going to be at ISMIR @ISMIRConf in SF from Nov 10-14! Looking forward to chat and discuss research with friends, old and new 🥳 #ISMIR2024

@0xhexhex Hi, we used two open source datasets collected by folks at @mtg_upf: Saraga and Hindustani Raga Recognition Dataset. Let me know how the tinkering goes!

@NithyaIsMe Fantastic! What datasets did you use? And are these available for research? I'd love to tinker with stuff like raga classification, etc. Thanks!

We built a hierarchical generative model to sing Hindustani vocal melodies 🎤! We will be presenting this work at ISMIR 2024! @ISMIRConf 📝Paper: arxiv.org/abs/2408.12658 💻Code: github.com/snnithya/GaMaD… 👩🏽‍💻Demo: huggingface.co/spaces/snnithy… 🎧Samples: snnithya.github.io/gamadhani-samp…

Work done with the dream team @maneeshadkr @wuyusongwys @huangcza @antoine_caillon!!

Intending to use this model for interactive human-AI generation, we present two possible use cases with our model: (1) primed generation and (2) coarse pitch conditioning. Learn more and play around with these interactions in our demo (huggingface.co/spaces/snnithy…)!

This looks so cool! Excited to play around with this soon 🥳 Just fyi, seems like this video on the website doesn't have audio enabled @pika_labs @demi_guo_

🚨Understanding In-Context Learning: 1. Pretrained LLMs can implement learning algorithms to learn from data in-context. 2. Transformers can encode multiple algorithms for the same task and use one based on context at inference time. 3. Attention-free models also exhibit ICL.

@the_smg97 @arvshank Yeah that makes sense! So are you saying more discrete changes in speed vs. the current continuous change or just increasing the wavelength of the sine wave to make the change in speed slower and thus less obvious? Or maybe both?

Today @snpranav and I got into an argument over how frustrating it would be for the speed of a song to constantly change as a function of time. To settle it, I decided to use Bespoke to control the speed of 'Spain' by Chick Corea based on a modified sine wave :)

Last week, @imnotfady and I built KarenAI for the @bitcmp hackathon. KarenAI - your personal Karen that can help you dispute your credit card transactions, get sh*t done with customer support, and let you speak with a manager at the touch of a button using Ai! 🧵👇