A Review of Glottal Waveform Analysis

Glottal inverse filtering is of potential use in a wide range of speech processing applications. As the process of voice production is, to a first order approximation, a source-filter process, then obtaining source and filter components provides for a flexible representation of the speech signal for use in processing applications. In certain applications the desire for accurate inverse filtering is more immediately obvious, e.g., in the assessment of laryngeal aspects of voice quality and for correlations between acoustics and vocal fold dynamics, the resonances of the vocal tract should firstly be removed. Similarly, for assessment of vocal performance, trained singers may wish to obtain quantitative data or feedback regarding their voice at the level of the larynx. In applications where the extracted glottal signal is not of primary interest in itself the goal of accurate glottal inverse filtering remains important. In a number of speech processing applications a flexible representation of the speech signal, e.g., harmonics plus noise modelling (HNM) [74] or sinusoidal modelling [65], is required to allow for efficient modification of the signal for speech enhancement, voice conversion or speech synthesis. In connected speech it is the glottal source (including the fundamental frequency) that changes under a time-varying vocal tract and hence an optimum representation should track glottal and filter changes. Another potential application of glottal inverse filtering is speech coding, either in a representation similar to HNM, for example (but incorporating a glottal source), or as in [3], [11], [19] applying coding strategies termed glottal excited linear prediction (GELP) which use glottal flow waveforms to replace the residual or random waveforms used in existing code excited linear prediction (CELP) codecs. In the studies cited the perceptual quality of the GELP codecs is similar to that of CELP. The speaker identification characteristics of glottal parameters have also recently undergone preliminary investigation [64] (note this is quite different from investigating the speaker identification characteristics of the linear prediction residual signal). Identification accuracy up to approximately 70% is reported using glottal parameters alone. Future studies employing explicit combinations of glottal and filter components may provide much higher identification rates. In addition, the more that is understood regarding glottal changes in connected speech