This repository holds the source code for my M.Sc. thesis work on procedural engine sounds using neural audio synthesis. The implementation uses TensorFlow and builds on DDSP. The method works by training models to reconstruct audio examples of recorded engine sounds from fundamental frequency (f0) signals corresponding to the engine speed (RPM). My report is available here and online supplementary material is available here.

The code is split into four modules:

1. data: Code related to data processing. Contains scripts for recording on-board diagnostics (OBD) data, processing audio and OBD recordings, and preparing .tfrecord files.
2. docs: Contains the online supplementary material of the report.
3. evaluation: Code for evaluating datasets, training processes and trained models.
4. models: Components of the models, including decoders, synthesizers and losses.