Critical Design of Neural Audio Synthesizers for Live Musical Performance
Franco Santiago Caspe
Supervisors: Mark Sandler - Andrew McPherson
Queen Mary University of London - Centre for Digital Music
Additional Material: Audio Examples and Videos


This website contains audio examples and videos, as well to relevant links to additional material referenced in the thesis.

Chapter 5: Tone Transfer with FM Synthesis

Audio Examples

You can access the audio examples at the corresponding website:

DDX7 paper website

Envelope Learning paper website

FM Tone Transfer Plugin

Download link: FM tone transfer Plugin

Demo Video

 

Chapter 6: Representation Learning for Low-Latency Interaction

BRAVE Audio Examples

The full collection of audio examples for the chapter is available at the paper website.

Timbre Transfer with BRAVE

We also present an excerpt of the audio examples, highlighting the reconstruction results of percussive and tonal sounds of RAVE and BRAVE models trained on the Filosax and Drumset datasets.

Models Trained on the Drumset Dataset
We encode and decode original excerpts of Drumset, Beatbox and Candombe datasets, using RAVE and BRAVE models trained on Drumset, and present the results on the following table.

Instrument Original RAVE (Trained on Drumset) BRAVE (Trained on Drumset)
Drumset
Beatbox
Candombe

Models Trained on the Filosax Dataset
We encode and decode original excerpts of the Filosax, Svoice and Viola datasets, using RAVE and BRAVE models trained on Filosax, and present the results on the following table.

Instrument Original RAVE (Trained on Filosax) BRAVE (Trained on Filosax)
Filosax
Svoice
Viola

BRAVE Audio Plugin and Demo Videos

This link includes both a plugin to run BRAVE models and a set of demo videos showing the capabilities of low-latency representation learning.

Minifusion Audio Plugin

Chapter 7: Interaction Design with Neural Audio

Here we present several audio recordings of different BRAVE training configurations, discussed in the thesis. Each yields different representations.

Rendered Transformations for different BRAVE Configurations

Sound Example Acoustic Guitar: Chord and Arpeggio Electric Guitar: Muted Strings
Original (Input)
FM Tone Transfer (BRASS)
BRAVE Original
BRAVE Two-Phase
BRAVE One-Phase
BRAVE Grafted (VCTK)

Guitar Playing with VCTK Model

Example of guitar playing with a model trained on continuous speech. Its encoder can then be used as a fixed representation to train other models in a “Grafted” approach.

Tonal and Percussive Rendering for Different Representations

The following table presents audio examples of transformations of different violin articulation excerpts, processed by models trained on the same data, but which learn different representations.

It seems that one-phase models, and models grafted with one-phase encoders, learn representations with a good capacity of tonal rendering, where a tonal input seems to “bleed” through the model (highlighted in blue). Models trained on a two-phase approach, or grafted with a two-phase encoder, seem to render atonal outputs even in presence of tonal inputs, a phenomenom through which the model seems to “hallucinate” atonal sounds in response to a tonal input (in orange).

Bowing behind the bridge Ordinario Jeté/Ricochet
Input
Animals Two-Phase
One-Phase
Grafted (one-phase)
Dice Two-Phase
Grafted (two-phase)
One-Phase
Music Two-Phase
One-Phase