This AI improviser learns the sounds you choose, then listens while you play

Imagine giving a computer a tiny repertoire: a few guitar notes, muted knocks on the wooden body and scrapes across the strings. Those sounds could define the world in which you improvise together. Choosing that world is already a musical decision.

That choice is central to Taylor Brook’s Tensed Computer Improviser, or TCI. An Organised Sound article published on 15 October examines the system; its abstract specifies a small audio corpus chosen by the user, with no pretrained component. Assembling the material becomes part of making the instrument.[2]

Brook has been developing the project since 2022. His documentation describes software that sorts and labels short grains of audio by their sound characteristics, then analyses an incoming performance and responds through different behaviours. The connection runs from recorded material to live listening and back out as sound.[1]

A collection becomes a playable possibility

The important choice begins before the performance. A carefully selected corpus can establish a sonic character: sustained textures, abrupt attacks, rough resonances or an unusual instrumental technique. In his earlier Journal SEAMUS essay, Brook argues that choosing and shaping such a collection is itself artistic work. A dataset can be composed with an intention.[3]

To organise sound, software needs descriptions it can compare. One family present in TCI’s repository is MFCC analysis. FluCoMa’s documentation explains that these coefficients compress a sound’s frequency spectrum into a short numerical description of its overall contour. They are useful for comparing timbre: the character that helps distinguish sounds beyond the note being played.[4][5]

That is a deliberately selective kind of listening. A representation useful for grouping acoustic textures need not carry lyrics, musical intention or a listener’s associations. In the guitar example, an analyst might care about the difference between a ringing tone and a scrape even when both were produced at almost the same place on a string.

Playing changes the conversation

TCI’s live input is consequential because its responses are informed by the signal arriving during performance. The source repository separates several analysis families, including pitch, chroma and spectral features, and contains a range of response modules. That organisation makes sound analysis and response design visible parts of the project.[1][5]

For a performer, the resulting creative question is how to play into those relationships. A gesture can become both an audible contribution and information for the computer. The player hears the response and chooses the next gesture. Even a familiar collection can support a different exchange as the musician changes timing, attack or register.

Brook’s SEAMUS essay welcomes the imprecision and glitches that small collections can produce as material for exploration. What counts as a useful surprise is then a musical decision. An unexpected response can become something a performer answers, develops or abandons.[3]

Where to begin exploring

The project is available on GitHub, and Brook’s official page links an operation video and a build diary. It runs in Max; the author says a paid Max version is unnecessary. The documented integrations include FluCoMa, ml.star and GO.[1][5]

After an unproductive exchange, a player could revise the collection and try again: remove a sound that dominates, add a contrasting texture, or narrow the range. A repertoire small enough to know becomes something to practise with and gradually make one’s own.

Report an error or suggest a correction