Exploring Approaches to Multi-Task Automatic Synthesizer Programming

Read Paper See Code Get data

Video Presentation

Proposed Approaches

Two different approaches to multi-task automatic synthesizer pro- gramming are presented, and a baseline was constructed to compare these methods. We find that the joint-decoder approach performs best.

Joint Separate

Separate-Encoder Separate-Decoder (Baseline)

The baseline model uses multiple single task automatic synthesizer programming (ASP) models for a comparison to different multi-task approaches. The single task model is based on state of the art appraches to ASP based on Variational Autoencoder model[1][2].

Joint-Encoder Separate-Decoder

The Joint-Encoder Separate-Decoder approach uses n parameter decoders and n masks for each decoder to infer parameters for each synthesizer. The values of the mask are set during training to indicate which synthesizer was used to generate the ground truth audio signal. if they are all ones, the ground truth was generated by that synthesizer, so that decoder’s weights will be updated during training. if they are all zeroes then that decoder will be ignored during training. For inference, masks can be set to ones ore zeroes depending on the desired decoder’s output.

Joint Separate
Joint Separate

Joint-Encoder Joint-Decoder

The Joint-Encoder Joint-decoder approach attaches a single decoder to infer parameters. The dimensiality of the output parameters is the same for all synthesizer parameter vectors, so the parameter vector p is padded with zeros so that the size of the vector is always equal to size of the largest parameter vector. The dimensionality of all masks in this model are the same as well.

[1] Esling, Philippe, et al. "Flow synthesizer: Universal audio synthesizer control with normalizing flows." Applied Sciences 10.1 (2019): 302.
[2] Le Vaillant, Gwendal, Thierry Dutoit, and Sébastien Dekeyser. "Improving synthesizer programming from variational autoencoders latent space." 2021 24th International Conference on Digital Audio Effects (DAFx). IEEE, 2021.

Results

We use log spectral distance (LSD) between audio generated from ground truth parameters and predicted parameters to measure the performance of our models. The lower the LSD the better the model.

Model Mean LSD
Separate-Encoder Separate-Decoder (baseline) 86.8
Joint-Encoder Separate-Decoder 72.9
Joint-Encoder Joint-Decoder 74.2

The Joint-Encoder Separate-Decoder performs best because it has the advantage of being trained on all three synthesizers while having specialized decoders for each specific synthesizer's parameter vector.

Examples

Model Comparison

Ground Truth Audio Example S-S Audio Example J-S Audio Example J-J Audio Example

Performance on Different Timbre Qualities

Ground Truth Audio 100% harmonic Prediction Audio 100% harmonic Ground Truth Audio 20% harmonic Predction Audio 20% harmonic

Performance with Varying Latent Dimension Sizes

Ground Truth Audio Prediction (Latent Size 2) Predction (Latent Size 64)