Elias Varga Vienna

Music thatlistens back.

Composer and AI music researcher, Vienna

I write for small ensembles and build models that follow musicians as they play. In spring I teach the next cohort of Composing with Models.

Elias Varga at a grand piano on a dark concert stage, a laptop and two small speakers beside him
Scroll to play the first bars.
Programme noteAndante, with patience

I came to models through the rehearsal room.

I grew up in Szeged playing the viola badly and listening very well. In Vienna I studied composition, then spent four years teaching machines to hear the way players hear each other: the breath before an entry, the slight push of a bow before a change.

Since 2022 I work on my own, between commissions, research and teaching. Every model I build has to survive a rehearsal before it reaches a paper.

  • 2022 to nowComposer and researcher, independentCommissions, research and the course, Vienna
  • 2018 to 2022Research scientist, generative audio, TonfeldReal time accompaniment models
  • 2014 to 2018PhD, music informaticsQueen Mary University of London
  • 2008 to 2013MMus, compositionUniversity of Music and Performing Arts Vienna
Elias Varga in his studio in Vienna with a viola, a laptop and handwritten score pages
ResearchPoco a poco, always a beat behind

My research teaches models to follow, not lead.

Most music models are built to replace the player. Mine listen to a player and answer, a beat behind and never louder. They are small enough to run on a laptop on stage: Python and PyTorch for training, a Max patch for the concert.

38ms

Following without leading

Score following for live accompaniment. Median delay of 38 ms behind the player, used in 14 concerts so far.

Proceedings paper, 2024
41hours

Quartet Hours

String quartet takes recorded with paid players who agreed to research and teaching use. Access on request, no commercial training.

Open dataset, 2025
212musicians

Does it sound played?

Blind listening tests. The model's line passed as human 31% of the time, and the study lists where it failed: breath, bow changes and endings.

Listening study, 2026
Spectrogram, the model answering bars 9 to 12Drawn live in the page
WorksModerato, for players and one quiet machine

The pieces are where the research gets tested.

Each work asks one question of a model. In Antiphon the quartet plays a phrase and the model answers it, softer, from four small speakers on stage. One piece has no model at all, because not every piece needs one.

  • AntiphonString quartet and listening model14 min2025Premiered in Vienna
  • GlasshousePiano and generated room9 min2024Premiered in Graz
  • Small HoursVoice, five players and model18 min2023Premiered in Linz
  • WeatheringOrchestra, no model11 min2021Premiered in Budapest
A string quartet performing Antiphon on a dark stage with four small speakers on stands
Composing with ModelsThe course, ten weeks

Ten weeks, fourteen people, one concert.

For composers and producers who read or make music and want to write with a model as a partner. No coding needed: I prepare the patches, and anyone who codes gets the source. Taught in English. Two places in each cohort are free for students who could not otherwise come.

Spring, online
From 2 March
Fee
€640
Summer, Vienna
5 to 16 July
Fee
€1,180
When
Tuesdays, 18:00
Places
14 a cohort
  1. Weeks 1 to 3

    Listen like a model

    What a model hears in a recording, and what it misses.

  2. Weeks 4 to 6

    Write with one

    Constraints, prompts as scores, and editing what comes back.

  3. Weeks 7 to 9

    Build a listening part

    A part that follows your own piece in rehearsal.

  4. Week 10

    Play it

    A closing concert with the cohort, recorded for you.

Students around a piano and laptops in a high Viennese room during the course

Hold a place in the next cohort.

Pick a cohort and leave your name. Your place is held on this page and I write within a week with the reading list and the first patch.

Choose a cohort

Call and responseAd libitum

Ask a question. I answer in the next bar.

Your question is set as a bar on the stave, and the answer arrives as the next one. With the sound on, you hear the call and the reply.

Answers from Elias's notes on this page. Scripted replies, not Elias.Write to the real Elias
0 / 200