Hybrid Autoregressive Resonance Estimation and Density Mixture Formant Tracking Model

Author(s)

Ramirez, Miguel Arjona

• Document type: Journal article
• Source: IEEE ACCESS; v. 6, p. 30217-30224, 2018.
• Web of Science Citations: 0

Abstract

A novel formant tracker is proposed using the mixture models of t densities (tMMs) for vocal tract resonance frequencies estimated with a hybrid linear prediction (HLP) method. The hybrid integer-cycle pitch-synchronous linear prediction (LP) analysis improves the frequency resolution over voiced segments, leading to closer formant estimates than those provided by other LP methods. In conjunction with HLP, formant trajectories are shown to be more nearly tracked by tMMs than by Gaussian density models. Tests with synthetic voiced and whispered speech as well as with an annotated database confirm better performance than either tMM clustering after formant estimation based on different time-frequency representations or tracking after different LP methods. (AU)

• FAPESP's process: 15/25512-0 - Conditional Analysis of Audio and Speech Signals for Coding and Recognition
• Grantee: Miguel Arjona Ramírez
• Support Opportunities: Regular Research Grants