A set of audio features for the morphological description of vocal imitations

Enrico Marchetto; Geoffroy Peeters
DAFx-2015 - Trondheim
In our current project, vocal signal has to be used to drive sound synthesis. In order to study the mapping between voice and synthesis parameters, the inverse problem is first studied. A set of reference synthesizer sounds have been created and each sound has been imitated by a large number of people. Each reference synthesizer sound belongs to one of the six following morphological categories: “up”, “down”, “up/down”, “impulse”, “repetition”, “stable”. The goal of this paper is to study the automatic estimation of these morphological categories from the vocal imitations. We propose three approaches for this. A base-line system is first introduced. It uses standard audio descriptors as inputs for a continuous Hidden Markov Model (HMM) and provides an accuracy of 55.1%. To improve this, we propose a set of slope descriptors which, converted into symbols, are used as input for a discrete HMM. This system reaches 70.8% accuracy. The recognition performance has been further increased by developing specific compact audio descriptors that directly highlight the morphological aspects of sounds instead of relying on HMM. This system allows reaching the highest accuracy: 83.6%.