Automating manual user strategies for precise voice source analysis

A large part of the research carried out at the Phonetics and Speech Laboratory is concerned with the role of the voice source in the prosody of spoken language, including its linguistic and expressive dimensions. Due to the lack of robustness of automatic voice source analysis methods we have tended to use labour intensive methods which require pulse-by-pulse manual optimisation. This has affected the feasibility of conducting analysis on large volumes of data. To address this, a new method is proposed for automatic parameterisation of the deterministic component of the voice source by simulating the strategies used in the manual optimisation approach. The method involves a combination of exhaustive search, dynamic programming and optimisation methods, with settings derived from analysis of previous manual voice source analysis. A quantitative evaluation demonstrated clearly closer model parameter values to our reference values, compared with a standard time domain-based approach and a phase minimisation method. A complementary qualitative analysis illustrated broadly similar findings, in terms of voice source dynamics in various placements of focus, when using the proposed algorithm compared with a previous study which employed the manual optimisation approach.

► Review of state-of-the-art voice source parameterisation methods.
► New LF parameterisation method, using exhaustive search and dynamic programming.
► Quantitative evaluation novelly using reference values obtained from manual analysis.
► Quantitative and qualitative evaluations reveal comparably strong performance.