Two Sonification Methods for the MindCube

Don D. Haddad; Fangzheng Liu; Joseph A. Paradiso; Lancelot Blanchard

REVIEW 3 major objections 5 minor 28 references

Two Sonification Methods for the MindCube

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A handheld fidget cube drives real-time generative music in two mappings.

desk verdict A solid NIME system report pairing MindCube with RAVE latent diffusion; the sensor-to-RMS conditioning proxy is asserted without calibration or evaluation, but the paper honestly frames the emotion-regulation goal as future work. read the letter →

arxiv 2506.18196 v1 pith:PRXN27XR submitted 2025-06-22 cs.HC cs.AIcs.SDeess.AS

Fangzheng Liu , Lancelot Blanchard , Don D. Haddad , Joseph A. Paradiso This is my paper · ORCID

classification cs.HCcs.AIcs.SDeess.AS

keywords MindCubemusiccontrollerlatentdiffusionRAVEclassifier-freeguidancesonificationemotionregulationVCVRack

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the MindCube, a 3.3 cm handheld fidget-cube device with buttons, a joystick, a rolling disk, and a 9-DoF IMU, can work as an expressive real-time music controller. It demonstrates two mappings: a generative-AI mapping that conditions a latent diffusion model on a sensor-derived RMS energy proxy to produce calming or stimulating music, and a non-AI modular-synthesis mapping in VCV Rack. The authors argue the AI mapping matters because it turns a tangible controller into a way to navigate a meaningful latent space, with a measured generation latency around 1.05 seconds per chunk. The emotion-regulation motivation is explicit but not validated in this paper.

What carries the argument

The load-bearing object is the RMS conditioning formula $RMS_{cond}$, a weighted standard deviation of the sixteen sensor signals over a moving window, normalized by $R$ and used as the conditioning value $c$ in the score function $\nabla_z \log p_\theta(z|c) = (1-\gamma)\nabla_z \log p_\theta(z) + \gamma \nabla_z \log p_\theta(z|c)$. It is what connects user activity to the generative model: high sensor variance yields one RMS condition, low variance yields another, and the diffusion model uses this to generate high- or low-energy audio. The other central components are the RAVE variational autoencoder for faster-than-real-time decoding, Latent Diffusion with 512-length latents and 30 diffusion steps, and outpainting of the previous latent tail for smooth continuation.

What would settle it

Record a user's interaction with the MindCube while simultaneously capturing ground-truth audio loudness or a physiological stress measure across varied manipulation patterns; if $RMS_{cond}$ does not correlate with the intended energy class, or if high-activity periods do not correspond to heightened stress, the AI mapping's responsiveness and its emotion-regulation motivation fail. The paper itself notes that emotion detection is not formally proven, so a direct correlation study would settle it.

Watch

Extended reading notes

Core claim

The central discovery is a working pipeline from physical fidgeting to generated audio: sensor streams are summarized into a weighted standard deviation $RMS_{cond}=(1/R)\sum_{i=1}^{16} w_i \sigma_i$, normalized to the range used during training, and used as the conditioning variable in Classifier-Free Guidance for a Latent Diffusion Model; the resulting latent codes are decoded by RAVE into 23-second audio chunks with about 1.05 seconds of generation latency. The paper describes this as a way to infuse meaning within a latent space: the sensor activity injects a musical-energy meaning into an otherwise abstract latent traversal. It also implements a second mapping that sends pitch, roll, joystick, button, and encoder values into VCV Rack to control filters, LFOs, panning, and gates, showing that the same hardware can drive both generative and handcrafted synthesis.

Load-bearing premise

Everything rests on the assumption that the weighted standard deviation of the MindCube's sensor readings is a valid real-time proxy for how active the user is, and ultimately for their emotional state; the paper states this is observed empirically but presents no measurements.

Editorial extensions

If this is right

The MindCube can serve as a portable one-hand music controller despite its 3.3 cm form factor, with BLE streaming at 20 Hz.
Sensor-activity-derived RMS can steer latent diffusion to produce contrasting high- and low-energy audio, making the latent space navigable by an external physical controller.
With about 1.05 seconds per generation, the system can read input roughly once per second and still feel responsive.
The same hardware supports a non-AI expressive mapping into VCV Rack, where pitch and roll, joystick, buttons, and encoder drive synthesis parameters.
If the activity-to-emotion hypothesis holds, the system could deliver calming music during high activity and stimulating music during low activity, supporting emotion regulation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The RMS_cond proxy could be tested independently of emotion: a studio recording with known loudness could verify whether the sensor-derived RMS actually tracks intended audio energy across different manipulation styles; the paper does not report such an evaluation.
The same conditioning mechanism could generalize to other continuous control signals, such as heart rate or breathing, since the pipeline only needs a scalar normalized condition.
Latency could be improved by chunk pipelining or by reducing diffusion steps, suggesting a concrete path toward tighter real-time coupling than the current 1.05 seconds.
The lack of user studies means the emotion-regulation claim is a design hypothesis; a controlled comparison of self-reported affect before and after use would be the natural next step.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two Sonification Methods for the MindCube is a NIME-style system paper, and the honest way to read it is as a carefully described integration rather than a hypothesis test. The genuinely new piece is the pairing: a small tangible fidget cube with a RAVE latent diffusion model, conditioned on RMS via classifier-free guidance, plus a handcrafted VCV Rack mapping. The outpainting continuation trick to keep generated audio flowing is a nice practical detail, and the reported 1.05s generation latency on an M3 Max is concrete and useful.

The paper does what a good system report should: it gives enough hardware/firmware detail to reproduce (three PCBs, nRF52832, BLE at 20 Hz with COBS), and it ships the AI mapping code on GitHub. The authors are also candid about the limits—they explicitly say they are not proving emotion detection and aim to do that in future work. That honesty deserves credit.

Now the soft spots. The load-bearing condition of the AI mapping is the sensor-to-RMS proxy, RMS_cond = (1/R) * sum(w_i * sigma_i). The paper asserts that accelerometer, joystick, buttons, and encoder are the best indicators, but gives no data, no weight values, no normalisation factor, and no calibration to the training RMS range (0–0.8724). Without that, the CFG conditioning could be sending arbitrary signals, and the claimed contrast between high-activity/quiet music and low-activity/loud music is not demonstrated. There is no audio-level evaluation, no perceptual check, and no user study. These are real gaps for a paper whose abstract promises “a way to infuse meaning within a latent space.” The stress-test note is right to flag this as the crux.

However, I would not call this fatal. For a NIME system report, the standard is that the system is plausible and implemented, and the authors frame the emotion-regulation angle as a working hypothesis. The missing calibration can be fixed with a short measurement section and a small listening test; the paper would be stronger with that, but it is not incoherent without it. The VCV mapping is thinner—more of a sketch than a full treatment, and no code is released—but it is explicitly presented as a contrast.

Who is this for? The NIME community and anyone building tangible controllers for generative audio will find it a useful reference. It deserves a serious referee: a system paper with concrete implementation, honest limitation statements, and a falsifiable (if not yet falsified) core claim. I would send it to peer review and ask the authors to add calibration details and at least one sanity-check listening example. That is a reasonable revision path, not a rejection.

Referee Report

3 major / 5 minor

Summary. The paper presents the MindCube, a 3.3 cm handheld device with buttons, joystick, rolling disk, IMU, and haptic motor, and describes two sound-mapping methods. The first uses a RAVE autoencoder with a latent diffusion model conditioned on an RMS value derived from weighted sensor standard deviations (Eq. 1 in §3.4), aiming to generate high-energy music during low user activity and low-energy music during high activity for emotion regulation. The second maps sensor data through a Python TCP server to a custom VCV Rack module for modular synthesis control. The paper reports a generation latency of approximately 1.05 s per latent chunk and provides an open-source implementation, but it explicitly defers emotion-detection validation to future work.

Significance. The paper contributes a concrete hardware platform and an unusual integration of latent diffusion with classifier-free guidance using a sensor-derived conditioning signal, and it reports measured generation latency and open-source code, which are positive aspects. However, the central claim that the AI mapping 'infuse[s] meaning' and provides expressive real-time control is not validated: the sensor-to-RMS mapping is uncalibrated, and no audio-level or perceptual evaluation is provided. The non-AI VCV mapping is similarly unevaluated. If the conditioning path were calibrated and the audio output verified, the system would be a useful demonstration of real-time latent-space navigation with a tangible controller; in its current form, the contribution is primarily a system description.

major comments (3)

[§3.4, Eq. (1)] The conditioning signal is not calibrated. The formula RMS_cond = (1/R) Σ w_i σ_i uses unspecified weights w_i and a normalization factor R, and the text states that the result is 'normalized to fit between the conditioning values used during training' without giving the normalization procedure. Because the sensor-derived quantity has arbitrary units and the training RMS range is 0–0.8724, the mapping can easily produce out-of-range or saturated conditioning values. The assertion 'Empirically, we observe that the accelerometer, the joystick, the buttons, and the encoder are the best indicators' is presented without supporting data. Without a validated transfer function, the CFG conditioning in §3.2 may be effectively random, so the claim that the mapping 'infuse[s] meaning within a latent space' (Abstract) is not established.
[§3.2/§3.4 and §4] Neither sonification method is evaluated. For the AI mapping, the paper provides no objective check that generated audio matches the conditioning RMS value (e.g., measuring the output audio's RMS against c), no spectrograms or audio excerpts, no listening test, and no user study. For the VCV Rack mapping in §4, there is likewise no demonstration of latency, responsiveness, or musical utility. Consequently, the central claim that the MindCube serves as an expressive real-time music controller through these mappings is supported only by description, not evidence.
[§3.4] The real-time claim is underspecified. The paper reports a generation latency of about 1.05 s per latent sequence and states that the sensor input is read at a rate 'lower than 1/1.05≈0.9524 Hz', yet it also says the system 'reads the data from the input sensors every second'. Since a latent sequence of length 512 corresponds to roughly 23 s of audio at 44.1 kHz, the relation between the 1.05 s generation cadence and the audio playback timeline is not explained, and no end-to-end latency from user gesture to audible change is reported. Without this, the 'responsive interface' claim is not substantiated.

minor comments (5)

[§2.1] The text refers to 'The system diagram in Figure 5', but the system diagram appears as Figure 3; unless the figures are renumbered, the cross-reference should be corrected.
[§3.4] The statement 'we are forced to read the sensor input at a rate lower than 1/1.05≈0.9524 Hz' conflicts with the earlier sentence that the system reads sensor data every second (1 Hz > 0.9524 Hz); the intended sampling and control rate should be stated consistently.
[§4] The final paragraph contains a typo, 'patchinwg', which should be 'patching'.
[§3.3] The phrase 'we train the model over 177 epochs until we observe the validation score going up' is ambiguous; presumably the validation loss starts increasing, and the sentence should say so explicitly.
[§3.2] The conditioning value c is stated to be normalized to the range [0,1], while the training RMS range is reported as [0,0.8724]; the paper should clarify whether the conditioning embedding expects values in the training range or a rescaled range.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sensor-RMS proxy is an unvalidated design choice, not a fitted parameter disguised as a prediction.

full rationale

The paper reports a prototype integration (MindCube sensors to RAVE latent diffusion with classifier-free guidance on RMS, plus a VCV Rack mapping) and measured latency; there is no derivation chain in which an output is shown to equal an input by construction. The RMS_cond formula in Section 3.4 defines a conditioning scalar from weighted sensor standard deviations, with the weights asserted from 'Empirically, we observe...' rather than fitted to a target outcome. That is an unvalidated empirical proxy, and the paper explicitly disclaims emotion detection ('we do not formally prove here that data from the MindCube can detect a user's emotional status'). Such an uncalibrated proxy and the absence of audio-level evaluation are correctness or validity risks, not circularity. The only self-citation [19] describes the MindCube hardware and is not used to justify the mappings; the training corpus (FMA) and model architectures (RAVE, latent diffusion, CFG) are cited external work. The reported 1.05-second latency is a measurement, not a prediction produced by fitted parameters. Therefore no significant circularity is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The free parameters are the hand-chosen sensor weights and normalization in the RMS_cond formula, which directly control the AI mapping's behavior but are neither reported numerically nor cross-validated. The axioms include the assumed validity of RMS as a loudness proxy and the hypothesis that fidget activity reflects emotional state. The only invented entity is the MindCube itself, reused from prior work, with no independent evidence for its emotion-sensing capability.

free parameters (4)

Sensor weight vector w_i = not specified numerically
Chosen by hand to favor accelerometer, joystick, buttons, and encoder; used in the RMS_cond formula in Section 3.4.
Normalization factor R = not specified
Scales the weighted standard-deviation sum to the training RMS range [0, 0.8724].
CFG guidance weight gamma = not specified
Controls conditioning strength in the score function in Section 3.2; the value is not reported.
Diffusion steps = 30
Chosen as a latency/quality tradeoff; not justified by analysis.

assumptions (4)

domain assumption RMS is a valid proxy for perceived loudness and energy of music.
Used throughout Section 3 to condition generation; not validated perceptually.
domain assumption The FMA small subset (8,000 tracks, 8 genres) is an appropriate training distribution for the emotion-regulation music task.
Training is described in Section 3.3; no justification is given for this dataset choice.
ad hoc to paper Weighted standard deviation of sensor readings reflects user activity, which in turn indicates stress or relaxation.
Working hypothesis stated in Section 2.2: 'increased interaction with the MindCube may indicate heightened stress levels'; explicitly not proven.
domain assumption RAVE and latent diffusion architectures behave as described in their original papers.
The paper relies on [7] and [20] for the core generation quality, without re-validating them in this context.

invented entities (1)

MindCube hardware device
purpose: Tangible controller for music generation and emotion regulation.
The device is described in previous work [19]; this paper provides no validation of its emotion-sensing capability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Two Sonification Methods for the MindCube." pith.science (2026). https://pith.science/paper/PRXN27XR

@misc{pith2026250618196,
  author       = {Pith},
  title        = {Pith review of: Two Sonification Methods for the MindCube},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRXN27XR}},
  note         = {Machine review of arXiv:2506.18196}
}

read the original abstract

In this work, we explore the musical interface potential of the MindCube, an interactive device designed to study emotions. Embedding diverse sensors and input devices, this interface resembles a fidget cube toy commonly used to help users relieve their stress and anxiety. As such, it is a particularly well-suited controller for musical systems that aim to help with emotion regulation. In this regard, we present two different mappings for the MindCube, with and without AI. With our generative AI mapping, we propose a way to infuse meaning within a latent space and techniques to navigate through it with an external controller. We discuss our results and propose directions for future work.

Figures

Figures reproduced from arXiv: 2506.18196 by the authors.

**Figure 3.** The system diagram of the MindCube. The MindCube is powered by a 100 mAh Li-Po battery, providing up to more than three hours of battery life during continuous data transmission. Additionally, a linear vibration motor is mounted inside, which can be programmed to deliver various haptic feedback patterns. The motor is controlled via pulse-width modulation (PWM). 2.2 Firmware and communication The MindCube firmware … view at source ↗

**Figure 2.** Three PCBs inside a MindCube. The MindCube contains three PCB boards, each dedicated to a specific function. The main control board manages all control and communication processes. It is equipped with an nRF52832 (ARM Cortex-M4, Nordic) Bluetooth Low Energy (BLE) systemon-chip (SoC). Additionally, the board includes an ICM-20498 9-DoF IMU, which captures 3-axis accelerometer, gyroscope, and magnetometer data. This … view at source ↗

**Figure 4.** The architecture of the AI Music mapping for the [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Live VCV Rack Patch that connects to the Mind [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

[1]

Mattia Davide Amico, Luca Andrea Ludovico, et al . 2020. Kibo: A MIDI controller with a tangible user interface for music education. In Proceedings of the 12th International Conference on Computer Supported Education. 1: CSME . SCITEPRESS, 613–619

work page 2020
[2]

Kathleen B Aspiranti and David M Hulac. 2022. Using fidget spinners to improve on-task classroom behavior for students with ADHD. Behavior Analysis in Practice 15, 2 (2022), 454–465

work page 2022
[3]

Lindsey Biel. 2017. Fidget toys or focus tools. Autism File 74 (2017), 12–13

work page 2017
[4]

Adrien Bitton, Philippe Esling, Antoine Caillon, and Martin Fouilleul. 2019. Assisted Sound Sample Generation with Musical Conditioning in Adversarial Auto-Encoders. In Proceedings of the 22nd International Conference on Digital Audio Effects (DAFx-19),. Birmingham, UK

work page 2019
[5]

Igor Bonifacic. 2020. Noise Machine is a tiny MIDI controller for creating music on the go. https://www.engadget.com/noise-machine-midi-controller- 231326364.html

work page 2020
[6]

Mason Bretan, Sageev Oore, Jesse Engel, Douglas Eck, and Larry Heck. 2017. Deep Music: Towards Musical Dialogue. InProceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI’17) . AAAI Press, 5081–5082. Place: San Francisco, California, USA

work page 2017
[7]

Antoine Caillon and Philippe Esling. 2021. RAVE: A variational autoencoder for fast and high-quality neural audio synthesis. arXiv:2111.05011 [cs.LG] https://arxiv.org/abs/2111.05011

arXiv 2021
[8]

Guilherme Campos, Nuno Fonseca, Anibal Ferreira, and Matthew Davies

work page

Show all 28 references

[9]

Nutan Chen, Djalel Benbouzid, Francesco Ferroni, Mathis Nitschke, Luciano Pinna, and Patrick van der Smagt. 2022. Flat Latent Manifolds for Human- machine Co-creation of Music. https://arxiv.org/abs/2202.12243 _eprint: 2202.12243

2022 arXiv
[10]

Patrick Chwalek and Joe A Paradiso. 2019. CD-Synth: a Rotating, Untethered, Digital Synthesizer.. In NIME. 371–374

2019
[11]

Suzanne B da Câmara, Rakshit Agrawal, and Katherine Isbister. 2018. Iden- tifying children’s fidget object preferences: toward exploring the impacts of fidgeting and fidget-friendly tangibles. In Proceedings of the 2018 Designing Interactive Systems Conference. 301–311

2018
[12]

Matson Driesen, Joske Rijmen, An-Katrien Hulsbosch, Marina Danckaerts, Jan R Wiersema, and Saskia Van der Oord. 2023. Tools or Toys? The Effect of Fidget Spinners and Bouncy Bands on the Academic Performance in Children With Varying ADHD-Symptomatology. Contemporary Educationa...

2023
[13]

Rebecca Fiebrink, Dan Trueman, and Perry R. Cook. 2009. A Meta-Instrument for Interactive, On-the-Fly Machine Learning. In New Interfaces for Musical Expression. https://api.semanticscholar.org/CorpusID:9059668

2009
[14]

Jules Françoise. 2013. Gesture–sound mapping by demonstration in interactive music systems. In Proceedings of the 21st ACM International Conference on Mul- timedia (Barcelona, Spain) (MM ’13). Association for Computing Machinery, New York, NY, USA, 1051–1054. https://doi.org/1...

2013
[15]

Ohad Fried and Rebecca Fiebrink. 2013. Cross-modal Sound Mapping Using Deep Learning. In Proceedings of the International Conference on New Inter- faces for Musical Expression . Graduate School of Culture Technology, KAIST, Daejeon, Republic of Korea, 531–534. https://doi.org/...

2013 doi
[16]

Michael Gurevich and Stephan von Muehlen. 2020. The Accordiatron: A MIDI controller for interactive music. arXiv preprint arXiv:2010.01574 (2020)

2020 arXiv
[17]

Jonathan Ho and Tim Salimans. 2021. Classifier-Free Diffusion Guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applica- tions. https://openreview.net/forum?id=qw8AKxfYbI

2021
[18]

Arnold, Brenton Par- tridge, Josiah W

Cheng-Zhi Anna Huang, David Duvenaud, Kenneth C. Arnold, Brenton Par- tridge, Josiah W. Oberholtzer, and Krzysztof Z. Gajos. 2014. Active learning of intuitive control knobs for synthesizers using gaussian processes. In Proceed- ings of the 19th International Conference on Int...

2014
[19]

Fangzheng Liu, Don Derek Haddad, and Joe Paradiso. 2024. MindCube: an Interactive Device for Gauging Emotions. In Adjunct Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology . 1–2

2024
[20]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

2022
[21]

Hugo Scurto and Ludmila Postel. 2023. Soundwalking Deep Latent Spaces. In Proceedings of the International Conference on New Interfaces for Musical Expression, Miguel Ortiz and Adnan Marquez-Borbon (Eds.). Mexico City, Mexico, 232–235. https://doi.org/10.5281/zenodo.11189166 I...

2023 doi
[22]

Victor Shepardson and Thor Magnusson. 2023. The Living Looper: Rethinking the Musical Loop as a Machine Action-Perception Loop. In Proceedings of the International Conference on New Interfaces for Musical Expression , Miguel Ortiz and Adnan Marquez-Borbon (Eds.). Mexico City, ...

2023 doi
[23]

Giuseppe Torre, Kristina Andersen, and Frank Baldé. 2016. The Hands: The making of a digital musical instrument. Computer Music Journal 40, 2 (2016), 22–34

2016
[24]

Sam Trolland, Alon Ilsar, Ciaran Frame, Jon McCormack, and Elliott Wilson
[25]

Elaine L Wong, Wilson YF Yuen, and Clifford ST Choy. 2008. Designing wii controller: a powerful musical instrument in an interactive music performance system. In Proceedings of the 6th International Conference on Advances in Mobile Computing and Multimedia. 82–87

2008
[26]

Kieran Woodward and Eiman Kanjo. 2020. ifidgetcube: Tangible fidgeting interfaces (tfis) to monitor and improve mental wellbeing.IEEE Sensors Journal 21, 13 (2020), 14300–14307

2020
[2018]

In Proceedings of the 21st International Conference on Digital Audio Effects (DAFx-18),

Generative Timbre Spaces: Regularizing Variational Auto-Encoders with perceptual Metrics. In Proceedings of the 21st International Conference on Digital Audio Effects (DAFx-18),. Aveiro, Portugal
[2022]

In NIME 2022

AirSticks 2.0: Instrument design for expressive gestural interaction. In NIME 2022. PubPub

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Mattia Davide Amico, Luca Andrea Ludovico, et al . 2020. Kibo: A MIDI controller with a tangible user interface for music education. In Proceedings of the 12th International Conference on Computer Supported Education. 1: CSME . SCITEPRESS, 613–619

work page 2020

[2] [2]

Kathleen B Aspiranti and David M Hulac. 2022. Using fidget spinners to improve on-task classroom behavior for students with ADHD. Behavior Analysis in Practice 15, 2 (2022), 454–465

work page 2022

[3] [3]

Lindsey Biel. 2017. Fidget toys or focus tools. Autism File 74 (2017), 12–13

work page 2017

[4] [4]

Adrien Bitton, Philippe Esling, Antoine Caillon, and Martin Fouilleul. 2019. Assisted Sound Sample Generation with Musical Conditioning in Adversarial Auto-Encoders. In Proceedings of the 22nd International Conference on Digital Audio Effects (DAFx-19),. Birmingham, UK

work page 2019

[5] [5]

Igor Bonifacic. 2020. Noise Machine is a tiny MIDI controller for creating music on the go. https://www.engadget.com/noise-machine-midi-controller- 231326364.html

work page 2020

[6] [6]

Mason Bretan, Sageev Oore, Jesse Engel, Douglas Eck, and Larry Heck. 2017. Deep Music: Towards Musical Dialogue. InProceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI’17) . AAAI Press, 5081–5082. Place: San Francisco, California, USA

work page 2017

[7] [7]

Antoine Caillon and Philippe Esling. 2021. RAVE: A variational autoencoder for fast and high-quality neural audio synthesis. arXiv:2111.05011 [cs.LG] https://arxiv.org/abs/2111.05011

arXiv 2021

[8] [8]

Guilherme Campos, Nuno Fonseca, Anibal Ferreira, and Matthew Davies

work page