AI & Music Data Scientist @moises_ai | prev. @c4dm
Joined June 2014
- Tweets2.4K
- Following1.9K
- Followers2.9K
- Likes19.8K
Pinned Tweet
soundcloud.com/hel9000/espel…
For this year's edition of the @aisongcontest we explored the Studio in @moises_ai and paired human vocals with brutal Portuguese generated screams!
Pedro Sarmento retweeted
“Start a sweep worker. As runs finish, compare results and launch the next useful experiment. Use only the allocated GPUs and keep a record of every decision.”
Powered by Cortex: github.com/SonyCSLParis/cort…
Pedro Sarmento retweeted
Core contributions of MIDI-REPA (midi-repa.vaclis.net/) by my RA @vaclis_:
1. AMTFlow: instrument tagger + MuScriptor for transcription
2. JamendoMIDI: audio data with transcribed MIDI
3. MIDI-REPA: align DiT hidden states with MIDI
4. TTM trainable on an RTX 3090 in 1.5 days
1/4
Introducing MIDI-REPA: improving text-to-music generation through symbolic representation alignment.
Melody can be represented compactly in symbolic form, so why do models still learn it from audio alone?
Paper: arxiv.org/abs/2610.20967
Demo & Code: midi-repa.vaclis.net
Pedro Sarmento retweeted
Turn a simple piano demo into a full pop song! We introduce `Piano2Pop` (hhhhaura.github.io/piano2pop…), a pianoroll-conditioned flow-matching model trained on psuedo piano-pop pairs generated by MuScriptor and PicoGen. Work done by my students @hhhhaura__, @ZionHzh & @tanchihpin0517
Excited to share our work — Piano2Pop: Expanding Piano Demos into Pop Productions Without Aligned Supervision !!!
check out our demo page: hhhhaura.github.io/piano2pop…
Pedro Sarmento retweeted
Our paper: Equivariant Music Transformer has been accepted to ISMIR 2026 (Best Paper Award Nominee). We show that music transformers become less equivariant as they scale or train longer, suggesting they memorize each musically-shifted version of the same data separately. 🧵👇
Pedro Sarmento retweeted
1/12 Introducing Music Lenses!
Something cool you can do with our recent sparse-inversion work is prompt a dense music space for concept-conditional similarity. You want similar vocal character, rather than just similar songs? Well, now you can!
pliploop.github.io/sparse-st… 🧵
Happy to share our #ISMIR2026 paper on steerable music retrieval with sparse autoencoders 🎛️🎧
Dense retrieval is great at finding similar tracks,
But once you have a query, what if you want to modify one concept inside it?
More piano,
More intimate,
Less gritty.
1/n
Pedro Sarmento retweeted
Jazz dataset alert! 🚨 Introducing JSIR (Jazz Standards dataset for Information Retrieval), an audio-aligned dataset of jazz recordings containing beat/downbeat, chord, and structure annotations. Repository, demo, and preprint out now: github.com/MALerLab/JSIR
Pedro Sarmento retweeted
as a runner, moving to SF meant improvising a run (as I often did in London/Santa Monica) could send me up a small mountain ... so I built: flattensf.com
Pedro Sarmento retweeted
in our upcoming ISMIR paper, Learning Jazz Pianist Style with Cross-Attention Conditioning, we ask whether a transformer can learn to play in the style of twelve giants of jazz piano. here our model continues one prompt in the styles of four of my favorite pianists. sound on! 🔊
Pedro Sarmento retweeted
if true, this is suuuuper interesting
the journalist quotes Jack Brody (chief product officer of Suno)
Suno is moving to models trained on music it has licensed. This is a big win for musicians, and likely wouldn’t have happened without the lawsuits of the last couple of years.
But they say the new models were still trained on “Suno user data”. This is worrying. It sounds to me like they have trained the new models on Suno output from previous models - output made with the old models, which were trained on musicians’ work without permission.
They also say the new models are trained on “preference learnings from its earlier models”. The same applies - if this is user preference data on existing Suno tracks, then they have it thanks to models trained on musicians’ work without permission.
If that is what they’re doing, it shouldn’t be allowed IMO. AI companies should not only shut off models trained on people’s work without permission - they should not be allowed to use the outputs or data generated from those models.
axios.com/2026/09/09/suno-v6…
Pedro Sarmento retweeted
#ISMIR2026 Tutorial Spotlight: “Introduction to #MusicInformationRetrieval: Three Perspectives”.
Be part of the ISMIR 2026 tutorial program in Abu Dhabi and connect with the latest topics shaping the music information retrieval community.
🔗 Learn more: ismir2026.ismir.net/tutorial…
Pedro Sarmento retweeted
What the what?? I hope this goes commercial.
David Wieland, Jonas Roth, Christoph Studer, "PolyMap: A 64-Channel Polyphonic Guitar Pickup System" arxiv.org/abs/2608.27522
Pedro Sarmento retweeted
📣 #ISMIR2026: Call for Sponsors!
Connect with leading researchers, innovators, and industry professionals shaping the future of #musictechnology.
🔗 Learn more: ismir2026.ismir.net/sponsors…
Pedro Sarmento retweeted
We already knew that "less is more" in music production, but turns out it's also true for bandwidth extension 🎶
Super cool findings by my teammates from @UMG, and definitely worth reading if you're interested in latent space geometry, music restoration or audio codecs 👀
New paper at #ISMIR2026! We discover that simple latent transport across multiple neural codec spaces can perform bandwidth extension "on par" with much larger models! How can we make use of this interesting latent geometry? Do we still need gigantic deep models for BWE?
Pedro Sarmento retweeted
Excited to share a new Preprint & LIVE DEMO! This is >2 years of my life building fast, lightweight CPU-friendly models to facilitate songwriters' iterative workflows, with Representations usable for Understanding and Generation, trained in a Self-Supervised way. First, the demo:
Pedro Sarmento retweeted
Call for Sponsors for the 27th International Society for Music Information Retrieval Conference (ISMIR 2026), taking place November 8–12, 2026.
Sponsoring ISMIR 2026 offers strategic value to your organization:
▪️Showcase Innovation
▪️Access Top Talent
ismir2026.ismir.net/sponsors…
Pedro Sarmento retweeted
Very proud to see our last work about Music Transcription released in open source. Go try it at muscriptor.kyutai.org
Inference code: github.com/muscriptor/muscri…
Paper: arxiv.org/abs/2607.08168v1
We're releasing MuScriptor, the best open model for multi-instrument transcription to date, created in collaboration with @MireloAI.
Give it a recording in any genre: pop, classical, metal, jazz, whatever, and it transcribes the individual instruments into MIDI. Link in 🧵
Pedro Sarmento retweeted
My thanks to Alex Ruger and TapeOp for a really fun interview- I enjoyed getting to talk about Roland Future Design Lab and my career with Roland, plus a trip down my own memory lane!
Behind The Gear with Roland's Paul McCabe tapeop.com/interviews/173/be… #tapeop via @tapeopmag
Pedro Sarmento retweeted
yeahh open source is fun 😜🤟
🥳 Announcing Stable Audio 3 🍕
🏆 fastest music models ever
💻 runs on MacBookPro M-series
🧪 break it plz
🧠 LoRA finetune in < 1h
📷 Sm = faster, Medium = qualityer
⚡ 59x realtime on M5 Pro
One-liner fast install:
curl -LsSf dadabots.com/_/sa3-mac | bash