From Sense to Sound Mapping High-Level Sonic Percepts to Sound Generation

This research report deals with the problem of mapping sonic percepts to sound generation, i.e. mapping high-level perceptually meaningful descriptors to sound synthesis parameters. We give an overview of different models that are used for sound description, sound generation and suggest a classification of the techniques that allow mappings from the descriptive to the generative domain. Based on our study of the state of the art, we describe our own novel system for intuitive control of an additive sound synthesis model. We in fact propose a general framework for the extraction, abstraction, reproduction and transformation of timbrical characteristics of a sound analyzed from recordings. We give a detailed description of our analysis/synthesis paradigm which is well-suited for a machine learning approach to a non-linear mapping problem. We also suggest an elegant way to reduce the dimensionality of the synthesis parameter space, while preserving the auditory quality of the resynthesis. We finally give a complete and coherent framework to train, tune and evaluate our system in an automatic, consistent and reproducible fashion. The results are promising, and this type of architecture yields various original audio and musical applications