Mission

Building the foundations of interactive music

Daydream is a music research lab developing models and real-time systems for human musical expression. We research how generative models can follow musical intention, respond to changing direction, and give creators precise control over sound as it unfolds.

Try previews Read the research

Possibilities

What if making music were as universal as listening to it?

When people can shape music in real time with precise, intuitive control, musical expression becomes part of everyday life. We see this research enabling new digital and physical instruments, new ways to develop and share musical ideas, and new forms of entertainment for music creators and their fans alike.

01
Musical instruments
Direct and perform with generative sound.
02
Creative tools
Develop musical ideas through continuous experimentation.
03
Interactive entertainment
Shape music through play, movement, and exploration.

Research

Precise musical control in real time

Our research asks how generative audio models can reliably follow chords, melodies, rhythms, and performance cues while the music is playing.

We develop specialized control models and the real-time systems that make this interaction possible. We evaluate how accurately models follow musical direction, how independently different controls operate, and how quickly the sound responds.

Live: Our real-time engine updates output sound within 100ms of a parameter change.
Additive: Each model is a small, modular add-on designed to steer any foundation model.
Musical: We’ve taught hallucination-prone models the language of music: chords, grooves, melodies, and more.

Fig. 1

Benchmarks

We benchmark our models using independent tools: chord detectors, pitch tracker, onset and transient detection, and more. The control is a leading open-source model.

TaskScoreControlOurs

Latency
Respond to a parameter change
Milliseconds from a parameter change to changed sound. Shorter is better · Control here is the closest known competitor.
Control200ms
Ours81ms
Chords
Play the written chart
Root accuracy against the chart, from an independent chord recogniser · A shuffled chart scores 11%.
Control9%
Ours69%
Melody
Play the line it’s given
Pitch-class accuracy, frame by frame, from a pitch tracker · The best published result we tested scores 48%.
Control15%
Ours53%
Groove
Land on the reference hits
Accent F1 against the reference groove · The wrong groove scores 37%.
Control59%
Ours80%
Accompaniment
Play the missing bass part
Onset match to the real part it replaces · Our hardest task. A generic input scores 25%.
Control11%
Ours33%
Mix
Mix toward a human engineer
Share of songs where ours sits closer to the human mix than the rough does
Control—
Ours79%
Note Density
Play more or fewer notes
Share of results where density moved in the correct direction
Control—
Ours100%

The engine

DEMON, our real-time audio engine, is open source. It answers a parameter change in under 100ms and takes control at 25 frames a second. Our models load into it as extensions.

DEMON on GitHub

Labs