Research / engineering / creative practice
Sound is a hard problem. Good.
Music technology is where signal processing, machine learning, hardware and human creativity collide. The Lab is where MML members learn those systems by building them.
We work in small project teams. Some questions come from members; others come from companies, university chairs or artists. Outputs can be prototypes, papers, performances or tools.
- 01
Represent
How should a machine encode a song: audio samples, notes, text, embeddings, or several views at once? Representation decides what a model can understand.
- 02
Generate
How can a system create seconds of convincing sound while keeping rhythm, harmony and form coherent across an entire piece?
- 03
Retrieve
Search, transcription, source separation, tagging and similarity often create more immediate value than generation.
- 04
Engineer
Classical DSP, real-time systems and physical hardware still matter. Our scope reaches from models to guitar pedals and chips.
- 05
Question
Who controls the tool, what data shaped it, and whose creative intent stays visible? Responsible practice is a design problem.
The real scale
One song is millions of decisions.
At 44.1 kHz, one second of audio contains 44,100 samples. A three-minute recording is almost eight million values before a system has represented timbre, rhythm, harmony or form.
Musical structure also lives at several timescales simultaneously: milliseconds shape timbre, seconds shape notes, and minutes shape the song. A useful model has to stay coherent at all of them.
That makes audio a valuable test bed for efficient sequence models, multimodal learning and new architectures, not merely another content format.
- 01 second
- 44,100
- 03 minutes
- ≈ 8M
- Timescales
- MS → MIN
raw samples
raw values
timbre to form
From the bench
One signal,
many scales.
One current direction explores wavelets and other multi-resolution function families. Music is hierarchical by nature; a transform that can zoom may fit it better than one fixed view.
The signal
one line, three scales at once
Coarse
song form
Mid
phrases & notes
Fine
texture & timbre
Fixed resolution · Fourier / STFT
Multi-resolution · wavelets
How work moves
Question → team → evidence.
The Lab is practical by design. Members learn methods while making something specific enough to test, share and improve.
01 / Frame
Make the question concrete
Define the musical need, technical constraint and evidence that would count as progress.
02 / Build
Prototype in a small team
Combine complementary skills, document decisions and get something working early.
03 / Share
Put it in front of people
A demo, paper, workshop or performance turns private learning into useful knowledge.
In good company
Partner network →
Have a question worth building?