A Spectral Glottal Flow Model for Source-filter Separation of Speech

Abstract

The estimation of glottal flow from a speech waveform is a key method for speech analysis and parameterization. Significant research effort has been made to dissociate the first vocal tract resonance from the glottal formant (the low-frequency resonance describing the open-phase of the vocal fold vibration). However few methods cope with estimation of high-frequency spectral tilt to describe the return-phase of the vocal fold vibration, which is crucial to the perception of vocal effort. This paper proposes an improved version of the well-known Iterative Adaptive Inverse Filtering (IAIF) called GFM-IAIF. GFM-IAIF includes a full spectral model of the glottis that incorporates both glottal formant and spectral tilt features. Comparisons with the standard IAIF method show that while GFM-IAIF maintains good performance on vocal tract removal, it significantly improves the perceptive timbral variations associated to vocal effort. * O. Perrotin and I.V. McLoughlin are with the

Glottal Flow Synthesis for Whisper-to-Speech Conversion

“…This begins with whisper input at the top right. The GFM-IAIF method [42] (Glottal Flow Model-based Iterative Adaptive Inverse Filtering, explained in section II-B) extracts the VT and glottis filters to be used in the whisper decomposition process in the top left panel. Successive inverse filtering of the vocal tract and glottis provides the GFD and the excitation, respectively.…”

“…Conversely, other methods such as Iterative Adaptive Inverse Filtering (IAIF) [49] focus on estimating both glottis and VT filter frequency responses from the signal regardless of its harmonicity. A variant called Glottal Flow Model (GFM)-IAIF [42] has been recently developed Glottis spectral envelope Fig. 3.…”

“…1. Compared to IAIF, GFM-IAIF has been shown to more accurately extract glottal formant and spectral tilt features [42], key parameters for distinguishing whisper from speech glottis filters. Figure 3 shows an example of speech and whisper decomposition.…”

“…The authors previously evaluated GFM-IAIF on static phones [42], but the current paper is its first use with continuous speech and whispers. Thus, we need to demonstrate that GFM-IAIF correctly decomposes the glottis and VT elements for those signals -specifically that the glottal parameters account for the main difference between V and UV versions of the same phone, while VT parameters remain broadly similar.…”

Hider-Finder-Combiner: An Adversarial Architecture for General Speech Signal Modification

“…Nevertheless, they are limited to controlling parameters which can be mapped to/from those speech features. GFM-IAIF [13] represents the spectral envelope with vocal tract and glottis-related parameters, for greater controllability [14]. Nonparametric methods avoid decomposition of the signal but are generally restricted to modifying one parameter.…”