Mateo Cámara

Parameter Optimization for a Physical Model of the Vocal System

Authors: Mateo Cámara, José Luis Blanco and Joshua D. Reiss

Abstract

We evaluated black-box and grey-box optimization techniques on the parameters of a simplified physical synthesizer, the Pink Trombone, to emulate both male and female vocal tract characteristics for vocalic and non-speech sounds. Leveraging prior research, we utilized a genetic optimizer and Mel-spectrogram representations to infer the articulatory parameters from human recordings through a direct spectral comparison with the synthesizer’s output. Optimization was carried out over temporal windows, introducing variations to the state-of-the-art objective metric to ensure temporal coherence across the synthesized signal. We also investigated grey-box approaches, utilizing algorithms such as pYIN for fundamental frequency prediction and developing a ResNet-based neural network as a foundation for the optimization process. Subjective tests validated our methodology and confirmed that the synthesizer effectively mimics human vocal sounds. Results demonstrate a marked superiority over previous state-of-the-art techniques, proving its practical utility and accuracy in real-world conditions. Moreover, these subjective evaluations allowed us to fine-tune the established perceptual metric ViSQOL, providing a calibrated tool for future research to assess auditory metrics in the domain of physical synthesizer modeling.

Video

ViSQOL Pink Trombone SVR Model

To use the Pink Trombone - ViSQOL model, first you should download the model:

Download SVR model

Next, you should check out the ViSQOL user guide. Once ViSQOL is installed, to use the model you should run the following code:

This will render as a formatted code block in the Markdown file:

./bazel-bin/visqol --reference_file ref1.wav --degraded_file deg1.wav --similarity_to_quality_model libsvm_svr_pinktrombone_model.txt

Link to the perceptual tests in Go Listen (in Spanish, use google translate if needed)

Link to Go Listen app

The questionnaire is available here

Pink Trombone online demo

A demo with all the yawns and some of the vowels is available online. This is just a tiny demonstration; all the audio samples used in the study are presented below.

Link to the interactive non-exhaustive demo