Effects of Training Data Variety in Generating Glottal Pulses from Acoustic Features with DNNs

Abstract

Glottal volume velocity waveform, the acoustical excitation of voiced speech, cannot be acquired through direct measurements in normal production of continuous speech. Glottal inverse filtering (GIF), however, can be used to estimate the glottal flow from recorded speech signals. Unfortunately, the usefulness of GIF algorithms is limited since they are sensitive to noise and call for high-quality recordings. Recently, efforts have been taken to expand the use of GIF by training deep neural networks (DNNs) to learn a statistical mapping between framelevel acoustic features and glottal pulses estimated by GIF. This framework has been successfully utilized in statistical speech synthesis in the form of the GlottDNN vocoder which uses a DNN to generate glottal pulses to be used as the synthesizer's excitation waveform. In this study, we investigate how the DNN-based generation of glottal pulses is affected by training data variety. The evaluation is done using both objective measures as well as subjective listening tests of synthetic speech. The results suggest that the performance of the glottal pulse generation with DNNs is affected particularly by how well the training corpus suits GIF: processing low-pitched male speech and sustained phonations shows better performance than processing high-pitched female voices or continuous speech.

A Comparison Between STRAIGHT, Glottal, and Sinusoidal Vocoding in Statistical Parametric Speech Synthesis

“…Based on this difference for GlottHMM, it can be argued that the shortcomings of the spectral model (from which GlottHMM suffers in analysis-synthesis) are averaged out in current SPSS acoustic models. Furthermore, when comparing the performance differences between GlottHMM and GlottDNN, we speculate that the quality of the DNNbased excitation of GlottDNN is highly voice-specific (for example, the performance of "Roger" decreased considerably because of the picked-up reverberation), as is also concluded in a separate study [64]. The simple excitation generation based on a high-quality glottal pulse utilized by GlottHMM is thus a safer option for more challenging voices.…”

Extraction and Utilization of Excitation Information of Speech: A Review

“…It is known that signal processing-based GIF methods are affected by distortions in the speech signal due to ambient noise, the poor audio quality of the recording equipment, and compression and bandwidth limitation caused by speech transmission [30], [269]. To address this issue, a few recent studies [269]- [271] have proposed using DNN-based methods for estimation of the glottal source waveform. In [269], coded telephone quality speech was studied using a DNN-based GIF method by using both clean and coded speech in training.…”