Automatic Selection of Synthesis Units from a Large Speech Database

Automatic Selection of Synthesis Units from A Large Speech Database Jau-Hung Chen and Chung-Hsien Wu* E000/CCL, Building 51, Industrial Technology Research Institute, Hsin-Chu, Taiwan, R.O.C. E-mail: chenjh@atc.ccl.itri.org.tw *Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, R.O.C. E-mail: chwu@csie.ncku.edu.tw ABSTRACT In this paper, a novel method for the selection of synthesis unit is proposed. The monosyllables are adopted as the basic synthesis units. A set of high-quality synthesis units is selected from a large continuous speech database based on four procedures: pitch period detection and smoothing, speech unit filtering, unit selection, and manual examination. Two cost functions are proposed for obtaining the synthesis units, which minimize the inter- and intra-syllable distortion. The cost functions estimate the parameters including the prosodic features, the LSP frequencies, and types of syllable concatenation. Experimental results showed that a match rate of 48.9% was achieved. It indicates that about half of the "best" synthesis units can be automatically obtained. Also, a replacement rate of 4.8% was obtained. 1. INTRODUCTION In past years, many studies have focused on TTS systems for different languages [1]-[2]. Also, TTS systems and synthesis technology for the Chinese language have been developed in the last two decades [3]-[4]. A detailed overview of TTS systems for English and Chinese were introduced by Klatt [1] and Shih and Sproat [5], respectively. Potential applications include aids for the handicapped, teaching aids, speech-to-speech translation, and any applications for text reading, such as email reader, news reader, and so on. General speaking, a TTS system could be logically composed of three main parts: text/linguistic analysis, prosodic information generation, and speech synthesis. Text analysis is first invoked to analyze the input text. The prosodic information generation employs the linguistic features to generate prosodic features including pitch contour, energy contour, and duration. Finally, speech synthesis is performed to modify the prosodic parameters of the synthesis units and generate intelligible and natural speech based on the above features. There are three modern approaches to speech synthesis: articulatory synthesis, formant synthesis, and concatenative synthesis [5]. The concatenative synthesis is the simplest and effective approach which uses real recorded speech as the synthesis units and concatenates them back together during synthesis. In concatenative speech synthesis, unit selection plays a prominent role of synthesizing intelligible, natural, and high-quality speech. In past years, many kinds of synthesis units have been proposed [1 ]. The phonemes have been adopted as the basic synthesis units. Such units take advantage of small storage. However, it needs to improve the accuracy of intra-syllable coarticulation and the spectral discontinuity between adjacent units. Consequently, longer synthesis units, such as diphone, demi-syllable, syllable, triphone and polyphone, are appropriately incorporated to reduce the effect of spectral distortion [2]. Recently, the approaches to unit selection from a large speech database or using non-uniform units [6] have been appreciated and proved to obtain natural and high quality speech. V Detection Method & Smoothing Method Filtering Criteria Two-Level Method) Synthesis Units Pitch Period Detection & Smoothing Spectral Feature Extraction (LSP …

Content truncated to 500 words