A fusion approach for automatic speech segmentation of large corpora with application to speech synthesis

Abstract

This paper deals with the automatic segmentation of large speech corpora in the case when the phonetic sequence corresponding to the speech signal is known. A direct and typical application is corpus-based Text-To-Speech (TTS) synthesis. We start by proposing a general approach for combining several segmentations produced by different algorithms. Then, we describe and analyse three automatic segmentation algorithms that will be used to evaluate our fusion approach. The first algorithm is segmentation by Hidden Markov Models (HMM). The second one, called refinement by boundary model, aims at improving the segmentation performed by HMM via a Gaussian Mixture Model (GMM) of each boundary. The third one is a slightly modified version of Brandt’s Generalized Likelihood Ratio (GLR) method; its goal is to detect signal discontinuities in the vicinity of the HMM boundaries. Objective performance measurements show that refinement by boundary model is the most accurate of the three algorithms in the sense that the estimated segmentation marks are the closest to the manual ones. When applied to the different output segmentations obtained by the three algorithms mentioned above, any of the fusion methods proposed in this paper is more accurate than refinement by boundary model. With respect to the corpora considered in this paper, the most accurate fusion method, called optimal fusion by soft supervision, reduces by 25.5%, 60% and 75%, the number of segmentation errors made by refinement by boundary model, standard HMM segmentation and Brandt’s GLR method, respectively. Subjective listening tests are carried out in the context of corpus-based speech synthesis. They show that the quality of the synthetic speech obtained when the speech corpus is segmented by optimal fusion by soft supervision approaches that obtained when the same corpus is manually segmented.