2. A genealogy of small-data machine improvisation

Interactive machine improvisation has, from its beginnings, favoured a small-data practice: its systems derive their behaviour from the individual performer (Chadabe Reference Chadabe1984; Rowe Reference Rowe1993; Pachet Reference Pachet2003), or learn from corpora curated by their makers (Schwarz Reference Schwarz2006, Reference Schwarz2007). For much of the tradition this smallness was a technical condition, a consequence of the computational limitations of real-time operation and of building systems that listen to a particular musician, rather than as an aesthetic commitment. The rise of large-scale generative models has thrown that condition into relief, and this retrospective view has revealed a difference in kind between models trained on material the maker has selected and heard versus models trained on corpora too vast for any individual to fully know. The first keeps the musician’s judgement in the loop at every stage; it is, in Pye’s terms, hospitable to a workmanship of risk.

George Lewis’s Voyager, whose first version was programmed between 1986 and 1988, remains a foundational example of interactive machine improvisation (Lewis Reference Lewis2000: 34), preceded by earlier real-time interactive systems such as Joel Chadabe’s Coordinated Electronic Music Studio and the League of Automatic Music Composers (Chadabe Reference Chadabe1984; Rowe Reference Rowe1993). Voyager did not learn in the statistical sense, as its ‘virtual improvising orchestra’ was rule-based, but it exemplified the two commitments the tradition has carried since: that an improvising machine performs with musicians interactively, and that its programme is a cultural text, representing ‘the particular ideas of their creators’ rather than any ‘objective’ or ‘universal’ musicality (Lewis Reference Lewis2000: 33). Statistical models of a performer’s own playing also entered this tradition in the mid-1980s, when Zicarelli’s Jam Factory (issued alongside the related programme M, by Chadabe’s company Intelligent Music) built Markov transition tables from a musician’s MIDI input and improvised upon them in real time (Zicarelli Reference Zicarelli1987). Pachet’s Continuator brought this approach to maturity, constructing variable-order Markov models of a musician’s input in real time and continuing their phrases in a stylistically consistent manner (Pachet Reference Pachet2003). In the factor-oracle lineage at Institute for Research and Coordination in Acoustics/Music, the MIDI-based OMax (Assayag et al. Reference Assayag, Bloch, Chemillier, Cont and Dubnov2006) was extended first to live audio through pitch-tracking into a symbolic representation, then to continuous audio descriptors by the Audio Oracle (Dubnov et al. Reference Dubnov, Assayag and Cont2007), and developed through a family of agents, the scenario-guided ImproteK (Nika et al. Reference Nika, Chemillier and Assayag2016) and the reactive SoMax2 (Borg et al. Reference Borg, Assayag and Malt2022) among them. What these systems train on typically amounts to minutes of music, such as a single session’s playing or a handful of recordings, and frequently the present performance itself; ImproteK’s memory, for instance, could be constituted ‘by recording the music played by the human co-improvisers during a live performance’ (Nika et al. Reference Nika, Chemillier and Assayag2016: 5).

In parallel, corpus-based concatenative synthesis established the audio corpus as a navigable space, organised by descriptor analysis and traversed by similarity (Schwarz Reference Schwarz2006, Reference Schwarz2007), and CatOracle joined the two strands, coupling CataRT’s corpus navigation to Audio Oracle pattern learning for improvisation in which ‘the user’s creative aural judgements are favored over theoretical criteria’ (Einbond et al. Reference Einbond, Schwarz, Borghesi and Schnell2016: 141).

The TCI descends from both currents at once. From the interactive systems lineage it takes the premise of a machine that performs interactively alongside a musician and that embodies its maker’s particular ideas rather than a neutral musicality; from concatenative synthesis it takes the audio corpus as a navigable space, traversed by descriptor similarity and sounded by replaying actual recorded segments of audio. What allows it to join the two is the arrival of machine-learning algorithms inside Max through the FluCoMa toolkit (Tremblay et al. Reference Tremblay, Roma and Green2021). The TCI thus brings together the improvising partner of the Voyager lineage and the corpus-as-instrument of concatenative synthesis: an interactive improviser whose entire musical world is a small, chosen body of recorded sound.

A second current in interactive machine improvisation makes the small-data commitment explicit: interactive machine learning as instrument building. The Wekinator demonstrated that models trained ‘on the fly’ from a handful of examples supplied by the musician could function as a meta-instrument (Fiebrink et al. Reference Fiebrink, Trueman and Cook2009), extending a use of machine learning in instrument design that Fiebrink and Sonami (Reference Fiebrink and Sonami2020: 237) trace back more than 25 years. The practice that followed, most extensively documented in Fiebrink’s eight-year collaboration with Laetitia Sonami on the Spring Spyre (Fiebrink and Sonami Reference Fiebrink and Sonami2020), has developed a working aesthetics in which the smallness and intimacy of training data are treated as creative affordances rather than statistical deficiencies (Fiebrink Reference Fiebrink2019; Vigliensoni and Fiebrink Reference Vigliensoni and Fiebrink2023). Young and Blackwell’s account of live algorithms supplies the tradition’s shared horizon: such research ‘is not concerned with systems that imitate human behavior; genuinely novel outcomes are sought, a product of renewed forms of human-computer interaction’ (Young and Blackwell Reference Young, Blackwell, Lewis and Piekut2016: 508). In this way, partnership and interaction become a site for expanding the boundaries of a musician’s aesthetic sensibility and expression (Brook Reference Brook2025).

The values stated by generative music research at scale are of a different kind, and the systems that state them deserve the same precision as the improvising lineage. The Music Transformer is presented as an advance in ‘generating music with long-term structure’, evaluated by perplexity and listener preference against a corpus of competition piano performances (Huang et al. Reference Huang, Vaswani, Uszkoreit, Simon, Hawthorne and Shazeer2019a); the Bach Doodle, which harmonised user melodies in the style of Bach for a mass audience, announces its achievement in its title as ‘approachable music composition with machine learning at scale’ and reckons that achievement in scale: more than 55 million harmonisation requests over three days, of which 21.6 million compositions, voluntarily contributed by their users, were released as a public dataset (Huang et al. Reference Huang, Hawthorne, Roberts, Dinculescu, Wexler, Hong and Howcroft2019b). The audio-domain systems that followed extend the trajectory from style modelling to end-to-end song production. A foundational model is trained once, by a research organisation, on a corpus assembled at a scale only such an organisation can hold, Jukebox on 1.2 million songs (Dhariwal et al. Reference Dhariwal, Jun, Payne, Kim, Radford and Sutskever2020), MusicLM on 280,000 hours of audio (Agostinelli et al. Reference Agostinelli, Denk, Borsos, Engel, Verzetti and Caillon2023), MusicGen on 20,000 hours of licensed music (Copet et al. Reference Copet, Kreuk, Gat, Remez, Kant, Synnaeve, Adi and Défossez2023), at a computational cost that places training beyond any individual practitioner. What a user then meets is the finished model’s conditioning interface: a text prompt, but increasingly also musical material, whether MusicGen’s melodic conditioning or MusicLM’s transformation of a hummed or whistled tune (Agostinelli et al. Reference Agostinelli, Denk, Borsos, Engel, Verzetti and Caillon2023; Copet et al. Reference Copet, Kreuk, Gat, Remez, Kant, Synnaeve, Adi and Défossez2023). The commercial song services, like Suno, built on the same foundation carry this interface to a mass public. Training determines the model’s sonic world and is complete before any musician arrives. These are substantial achievements on their own terms, but fidelity to a style distribution, coherence at scale and deployment to millions measure a workmanship of certainty, and they are values orthogonal to those that experimental traditions of improvised music have most prized. Practices as heterogeneous as those of the Association for the Advancement of Creative Musicians (AACM) and free jazz (Lewis Reference Lewis1996) and Pauline Oliveros’s deep listening (Oliveros Reference Oliveros2005) are affiliated less by a shared history than by shared values: not the faithful reproduction of a style but its creative disruption through the cultivation of difference in an improvisation that is an encounter with an unpredictable partner that Born, Lewis and Straw describe as ‘not only of social commentary but, potentially, of social experimentation’ (Born et al. Reference Born, Lewis and Straw2017: 9–10).

3. TCI technical overview

The TCI’s design follows the workflow its user encounters: a corpus of audio is imported, segmented and analysed; a bank of models is trained from those analyses; and in performance the software assembles its response by sampling short segments of the corpus with concatenative synthesis, steered by real-time analysis of the live input (Figure 1).

Figure 1.

Signal-flow diagram of the TCI. Build and analysis section: the corpus is reduced to a mono buffer, segmented in parallel into onset slices and grains of fixed duration, and measured along five descriptor families, yielding ten analyses that fit the model bank. Model bank section: for each analysis, a standardised–UMAP–grid pipeline produces 2D corpus maps, k-means yields cluster labels, a k-d tree provides a nearest-neighbour index and a variable-order Markov chain (maximum order three) is built upon the cluster labels and inter-onset durations. Performance section: live input is tracked with the five descriptors and queried against the k-d tree, whose matches and labels drive a layer of behaviours (see Figure 4); the segments they select are sounded by a 128-voice polyphonic sampler across a variable loudspeaker array, when the analysis is streamed over OSC to external processes such as live visuals as desired.

The interface presents the preparation of these models as a numbered sequence; its panels are labelled STEP 1 through STEP 4, and the improviser does not sound until the steps have been completed to train the models (Figure 2). Until a corpus is analysed, the descriptor maps are blank cells and its waveform display is empty. Importing and analysing an audio corpus is the act that allows the software to sound and function.

Figure 2.

The TCI’s main interface as it appears on launch, before a corpus has been analysed. The preparation of the models is laid out as a numbered sequence in the lower panel: STEP 1 (import audio corpus), STEP 2 (adjust settings), STEP 3 (onset detection) and STEP 4 (analysis). Until those steps are run, the descriptor maps remain as empty cells, the waveform display is empty and the TCI will produce no sound.

3.1. Building and segmenting the corpus

A TCI corpus may be a single audio file, a folder of files concatenated into one buffer, or audio recorded live from the input. The documentation advises keeping the source audio short to keep analysis and real-time matching within a workable latency budget.

The corpus is divided in two complementary ways. First, onset detection slices the audio buffer at points of attack or abrupt spectral change so that each slice is a sounding event of the kind a listener might parse as a single gesture. In parallel, the corpus is cut into grains of fixed duration, which capture continuous texture rather than discrete gesture. The two segmentations provide a gestural and a textural reading of the same material respectively. The threshold that governs onset detection is left in the user’s hands as a musical and interpretive decision, and the manual embedded in the software states that ‘the more active the onset detection is, the more saturated the resulting sound of the improviser will be’. Once detection has run, the interface displays the corpus waveform overlaid with its slice boundaries and reports the number of onsets found (Figure 3), making the consequences of the threshold immediately legible to the user.

Figure 3.

The TCI interface after a corpus has been imported and analysed. Each analysis is shown as a 2D corpus map aligning with the five descriptor families (MFCC, chroma, spectral, melbands and pitch) across the two segmentations (grains, top row; slices, bottom row). The points of the corpus maps are spread by the high minimum-distance UMAP setting and normalised to fit the plotters; the colours of the points represent the results of the unsupervised k-means clustering; the performer may click on a point on any map and hears the slice or grain of audio it represents. The corpus waveform and its slice boundaries appear at lower right, with the onset count reported beneath (here 1,918). Corpus shown: loadbang’s recording of the author’s Tarantism (loadbang 2021).

3.3. Listening, behaving, sounding

Once trained, the TCI listens to the live input with the same descriptors it used on the corpus, tracking the pitch, loudness, spectral profile and attacks of the audio input in real time, following the range of dynamics and estimating beat and metre. What it hears becomes a stream of queries: a live sound is matched, through the k-d trees, to its nearest neighbours in the corpus, so that the machine answers from its own material in terms set by the user.

How the TCI selects what samples or grains of audio it plays back is governed by a bank of 16 behaviours, grouped in three families. Many are offered as slices-or-grains or long-or-short pairs, and the reactive and generative behaviours can each be run on any of the five descriptor analyses, so that a single behaviour may sound in as many as five variants. Reactive behaviours track the live input, doubling its attacks with matched material and optionally sustaining them, echoing it granularly, following, inverting or detuning its amplitude envelope, harmonising to its chroma or deliberately contrasting its spectrum. Rhythmic behaviours impose or follow a pulse, locking corpus playback to the detected tempo and metre, matching pulse, or generating Euclidean patterns, ratchets, and accelerating ‘bouncing-ball’ and decelerating ‘slowing-down’ figures (built from the gen∼ abstractions of Wakefield and Taylor Reference Wakefield and Taylor2022). Generative behaviours draw on the corpus’s own statistics or on autonomous processes. The first is a variable-order Markov chain (ml.star; Smith and Garnett Reference Smith and Garnett2012), of maximum order three, trained on the sequence of cluster labels, plays out new labels in the order that this particular material tends to follow. The second generative behaviour is a Lorenz chaotic attractor that steers parameters along continuously unfolding trajectories. The third imposes a pulsation of grains or segments within a randomly selected label in from the model. This three-family grouping is analytic, organised by how each behaviour responds to the live input; in the patch itself the behaviours are laid out differently, first by segmentation (those acting on onset slices and those acting on grains) and then in rhythmic, textural, doubling and generative columns, alongside a separate set of one-shot gestures (the bouncing-ball and slowing-down figures among them) and the global effects (Figure 4). Table 1 regroups them by how they respond to the live input.

Figure 4.

The behaviour panel. The TCI’s behaviours are grouped first by segmentation (SLICES, top; GRAINS, below) and then into RHYTHMIC, TEXTURAL, DOUBLING and GENERATIVE columns, with a separate set of one-shot gestures (the bouncing-ball and slowing-down figures and a glitch drone) and the global effect parameters (dynamics, feedback delay, reverse delay) on the right. Behaviours change automatically during operation in reaction to the audio input. Additionally, behaviours can be turned on, monitored or blocked manually by the user. The article slightly regroups these modules analytically by how they respond to the live input as reactive/rhythmic/generative (Table 1).

Table 1.

Behaviour modules, grouped analytically by how they respond to the live input (Figure 4)

Table 1. Long description

A table categorizing audio playback behaviors into reactive, rhythmic, and generative groups with descriptions. The table has three main groups: Reactive, Rhythmic, and Generative, each with specific behaviors and descriptions. Reactive behaviors include Onset doubling, Doubling onsets and sustain, Granular matching, Envelope follower, Chroma harmonise, and Melband contrast. Rhythmic behaviors include BPM follower, Euclidean rhythms, Ratchet rhythms, Pulsation, Bouncing ball, and Slowing down. Generative behaviors include Markov continuous slices, Markov imitator, and Lorenz chaotic. Each behavior is described in terms of how it processes or responds to audio input.

Variants are shown in parentheses. Reactive and generative behaviours run on any of the five descriptor maps (pitch, melbands, spectral shape, chroma and MFCC) – the ‘up to five variants’ of each. The rhythmic generators (bouncing-ball, ratchet, Euclidean, pulsation) are built from the gen∼ abstractions of Wakefield and Taylor (Reference Wakefield and Taylor2022); the Markov behaviours use the ml.star MAX package (Smith and Garnett Reference Smith and Garnett2012). Grain sustain and glitch drone also appear in the behaviours window as toggles, but these are one-shot effects, not behaviours, and are excluded from the table.

In its fully automated mode the TCI supervises the behaviours itself, holding one or a few active at a time, switching among them by weighted chance, sensing whether a pulse is present, and forcing a change after five seconds of the performer’s silence. The behaviours are saturated with controlled randomness, so that no two passages through the same corpus are alike; the result is an improviser consistent in character but unpredictable in detail. A manual mode sets the supervisor aside and lets a performer audition or play individual behaviours directly.

As the TCI performs, it streams its analysis in the MAX console and via Open Sound Control (OSC) to local network ports so that the same listening that drives its playing can be made to drive external processes such as live visuals. The instrument that results is wholly contained in the patch and wholly the product of the material fed to it.

4. Small datasets as a workmanship of risk

Pye’s concepts provide an understanding about where in a process the quality of the outcome comes to be decided, and he observed that most real work mixes the two modes of risk and certainty. In industrial generative music the deciding act of the assembly and modelling of the training corpus happens once, before the system is ever played and on infrastructure the musician has no role in. A vast body of audio is gathered and a model trained on it at great computational expense by the developing organisation. The musician/user meets only the finished model, and the meeting is not confined to the text prompt; as Section 2 noted, contemporary systems accept melodic or audio conditioning that steers generation (Agostinelli et al. Reference Agostinelli, Denk, Borsos, Engel, Verzetti and Caillon2023; Copet et al. Reference Copet, Kreuk, Gat, Remez, Kant, Synnaeve, Adi and Défossez2023). But conditioning operates within a sonic world already fixed. The practitioner can neither intervene in the foundational training nor determine the corpus from which the model was built, and it is there, in training, that the aesthetic boundaries of the outcome are decided: the locus of Pye’s ‘risk’ has been engineered out of reach. The algorithms the TCI uses are deterministic procedures: given the same corpus and the same settings (with random seeds held fixed), UMAP, k-means and the rest return the same models every time. In Pye’s terms the apparatus fixes the result, which is the definition of the workmanship of certainty. However, the TCI puts the certainty of its algorithms back into the service of a workmanship of risk by returning control over what material the models are built from, how it is segmented, when and how often they are fitted.

The TCI’s models are not pretrained, but are built entirely from whatever audio is provided by the user. Choosing and recording the corpus are part of the creative act, reflecting the understanding that deciding which sounds are admitted to the corpus, and whose, is an aesthetic and social decision rather than a technical one (Born et al. Reference Born, Lewis and Straw2017). It is here, too, that the smallness of the corpus contributes to the aesthetic. The standardising tendency I have elsewhere identified as central to large-scale machine learning, the regression towards the statistical centre of a vast dataset (Brook Reference Brook2023a; Doshi and Hauser Reference Doshi and Hauser2024), finds insufficient data to act on when that dataset is ten minutes of a single musician’s playing. The distinction between training and conditioning is relevant; ten minutes of audio could not train a large generative model; and it could serve a pretrained model as conditioning or adaptation material, but material used that way steers a representation already derived from a much larger dataset, and the result remains answerable to that dataset’s statistical centre. In the TCI the same ten minutes is not a query put to a larger learned world; they constitute the model’s entire sonic world, and the boundary of the corpus is the boundary of what the TCI can sound. This bespoke fit of a corpus tailored to a single musician and invites a recognised objection; McCormack and colleagues observe, of real-time creative-AI systems generally, that they ‘tend towards being bespoke to their designer(s), making general or sustained creative development difficult’ (McCormack et al. Reference McCormack, Hutchings, Gifford, Yee-King, Llano and d’Inverno2020: 45). Read through Pye, the close fit of instrument to maker is not a deficiency to be engineered out but the condition of any craft tool, and the source of whatever identity the instrument comes to have.

Training a model from a small corpus of audio in the TCI fits in seconds to a few minutes, and inside the same patch that will perform, the division that structures industrial machine learning does not arise. Training is not infrastructure laid down before the work begins but folded into the work itself. Setting the onset threshold, choosing how finely the corpus is clustered, hearing the resulting map and re-analysing with a different setting are interpretive decisions that are reversible, repeatable and audible in their effects. This is the condition Fiebrink and Sonami describe from eight years of building and refining the Spring Spyre with machine learning: models shaped iteratively by a musician’s ear rather than optimised against a benchmark (Fiebrink and Sonami Reference Fiebrink and Sonami2020). It is what Pye names a workmanship of risk: a result whose quality depends on the maker’s judgement and care as the work proceeds. The effort this asks of the player is not friction to be designed away: as Michalakos, after Joel Ryan, observes, the effort an interface demands is closely bound to the expression it affords (Michalakos Reference Michalakos2021: 80). Effort alone does not make a craft; however, it earns its place where it exposes consequential decisions that become legible through practice. The labour of slicing, clustering and auditioning is of this kind, where every adjustment is audible in the maps and in the playing, and it is where the playing of this instrument begins.

A craftsperson knows their tools and materials intimately, and that knowledge is itself part of the skill (Sennett Reference Sennett2008). Understood this way, craft is less a category of objects than, in Adamson’s terms, a way of doing things, a process and a disposition rather than a class of artefact (Adamson Reference Adamson2019: 3–4). Radford gives that disposition its electroacoustic setting, describing the studio as an ecology of implements through which a practitioner’s skill and knowledge are mediated (Radford Reference Radford2015). The TCI reflects these values of craftsmanship by making the corpus maps the principal display of the user interface. The user sees how the machine has organised the corpus, can move through any map and audition its regions, and can therefore form an understanding of what the TCI is capable of. Where the prevailing rhetoric of AI trades on opacity, the ‘magic’ that Elish and boyd identify as obscuring how these systems work (Elish and boyd Reference Elish and boyd2018), the TCI instead makes its hearing clear and comprehensible.

Three further features follow from the corpus being audio rather than symbol. First, recorded sound carries the trace of the body that made it: in Iyer’s account, to hear music is to perceive the action of another agent (Iyer Reference Iyer, Lewis and Piekut2016: 87). What reaches the listener from the TCI can be heard as someone’s playing, not as a rendering of playing in general. Second is the technical condition on which the first depends; the TCI’s concatenative playback cuts, indexes and reassembles the corpus itself so that its output is always fragments of the source recordings rather than an approximation generated from a learned stylistic distribution. The embodied trace survives in the output because the TCI plays fragments of the corpus source audio, rather than newly generated. Third, a small corpus yields a sparse and uneven map; its clusters do not divide cleanly, and its nearest-neighbour matches are often approximate or relatively distant. These are the errors of data-poverty that can be creatively stimulating, and they are not engineered away in the TCI. These errors are where the instrument may exceed its maker’s intention, the surprise that McCormack and colleagues identify as a central value of generative systems, those able to ‘surprise and delight their authors’ (McCormack et al. Reference McCormack, Gifford and Hutchings2019: 39). The Scuffed Computer Improviser pursued such imprecision deliberately (Brook Reference Brook2023b); in the TCI, it arises of its own accord, from the smallness of the data, and is welcomed in the same spirit, as the tool-marks of a workmanship of risk. Unpredictability is not in itself a workmanship of risk; Pye’s term names the dependence of the result on judgement, dexterity and care, not the mere presence of uncertainty. As a performer develops the skill to anticipate the TCI behaviours, to provoke or avoid them, and to turn them to musical account, the workmanship of risk materialises.

Keeping machine learning small, then, moves the decisions that determine the music into the hands and ears of a musician working with chosen material in real time. The small-data approach also facilitates legibility regarding how the software is functioning and is a practice that can be shown and taught.

5. Making in public

Craft is never fully reducible to written instruction but passes from maker to maker by showing, watching and imitation. Indeed, Sennett places apprenticeship at the centre of craftsmanship (Sennett Reference Sennett2008). The TCI enters that tradition in two ways. First, it is open-source; the Max patch and its abstractions are published in full so that any user may open the instrument, read its construction and alter it. Second, the TCI was built publicly through a series of video streams documenting its making from scratch (Brook Reference Brook2023c). The video series exists as a record of the development of the TCI, in which the design decisions set out in Section 3 are shown being taken and worked through. What my previous arguments urged regarding the demystification of machine learning (Brook Reference Brook2023a, Reference Brook2025), the build series attempts to enact, making the instrument’s construction explained and reproducible.

The TCI is built on FluCoMa, a toolkit conceived to put corpus analysis and machine learning into musicians’ hands as a shared and documented practice rather than a proprietary service (Tremblay et al. Reference Tremblay, Roma and Green2021), and carried by an explicit pedagogy of teaching that accessibility (Moore et al. Reference Moore, Bradbury, Tremblay and Green2025), reflecting a culture of transmission and open-source access. This culture of transmission is seen within the patch itself as the TCI credits its sources where they are borrowed: the beat-and metre-detection derives from work contributed by Philippe Salembier via their Youtube comment; the chaotic and rhythmic generators are adapted from the gen∼ abstractions of Wakefield and Taylor (Reference Wakefield and Taylor2022); the learning objects are those of FluCoMa and of Smith’s ml.star library. These acknowledgements embody a model of authorship sharply opposed to the one prevailing in industrial generative AI, where the labour a system depends on is characteristically effaced; recordings absorbed into a training set return only as statistical tendency, a condition that has resulted in both ethical and legal difficulties (Morreale Reference Morreale2021). Two distinct failures are at work here, and they should be separated. The first is a lack of acknowledgement, and it is not universal; MusicGen was trained on licensed catalogues (Copet et al. Reference Copet, Kreuk, Gat, Remez, Kant, Synnaeve, Adi and Défossez2023), and Stable Audio Open is trained entirely on Creative Commons audio with its sources documented (Evans et al. Reference Evans, Parker, Carr, Zukowski, Taylor and Pons2024). The second is a lack of output-level traceability, and it survives even such disclosure; a licensed and documented corpus still returns as statistical tendency, and neither listener, developer nor contributing musician can trace a given output back to the recordings that shaped it. Nor does the effacement stop at musicians; the much-publicised sale of Edmond de Belamy as the work of an algorithm, when the network that produced it was in fact another developer’s, showed the same opacity operating on upstream technical labour: who built a system that can disappear as readily as whose recordings trained it (McCormack et al. Reference McCormack, Gifford and Hutchings2019: 35–7). The TCI’s chain of authorship is legible at both levels and at a finer grain: the corpus a musician has chosen and recorded, the thresholds they have set, the behaviours they have selected, and the borrowed components named in the patch are all acknowledged, open to inspection and audible in the output they shape.

For a demonstration of the TCI, I trained a corpus from Architek Percussion’s recording of my percussion quartet, Incantation (Architek Percussion 2018), whose bowed metal bars, brake drums and prepared hi-hats become the vocabulary from which the TCI improvises, its responses audibly made of that percussion writing even as they recombine it into figures the recording never held (Video example 1). The corpus in the video example is a single 8-minute recording of a percussion quartet being excited by a live performance using a mallet on a metal pipe. The mismatch of the live performance with corpus means that the TCI’s nearest-neighbour matches are frequently inexact, and one hears the consequence directly; a bright metallic attack in the live input is answered not by an identical or nearly identical sound but by its nearest available neighbour of a brake-drum stroke or a bowed bar of roughly similar brightness. With this interaction, the response shadows the input at a timbral distance, suggesting perceptual connections and variations rather than replica. Furthermore, the clusters drawn from so little material do not divide cleanly, and the variable-order Markov chain that plays out their labels strings the percussion into successions that are similar to the recording but never exactly the same, making audible a syntax that has been developed in the model. Where the corpus thins, the maps grow sparse and a handful of segments recur. A performer comes to recognise and, with familiarity, to solicit or sidestep these recurrences in the moment. These are the errors of data-poverty heard as tool-marks; the bowed bars, brake drums and prepared hi-hats remain recognisably themselves, their recorded bodies intact, even as the instrument recombines them into gestures that exceed what was ever played into it. A performer may accumulate, over time, a library of such corpora, each a particular sound-world to which the TCI can be returned.