Semiautomatic inverse filtering of vowels with differing degrees of nasality

Most approaches to inverse filtering are either manual or automatic. Manual methods, while providing high-quality fine-grained analysis, require user expertise and manpower to analyze even small amounts of data. Automatic methods allow large-scale analysis, but lack the precision required to capture non-neutral voice dynamics as present in, for example, expressive speech. This paper reports on development of a semiautomatic inverse-filtering toolkit, focusing on an adaptive nasal/non-nasal vowel inverse filter. For vowels with differing degrees of nasality (from not nasalized to heavily nasalized), each frame is analyzed using machine learning techniques to assess the degree of nasality. Based on this assessment, the inverse filter is adapted according to acoustic theory, to best analyze each frame of the data, employing either all-pole or pole-zero analysis where necessary. User intervention is possible, in both specifying analysis frames and fine-tuning of pole and zero number and location. Training data for machine learning are selected on the basis of auditory and acoustic analysis. This semiautomatic approach allows analysis of large-scale corpora while retaining the robustness and fine-grained analysis of manual methods. Results from analysing both non-nasalized and nasalized vowels from a corpus of Irish are presented. [This work is funded by the IRCSET Embark Initiative.]