Assessment Mode Of Chinese Speech Synthesis System

This paper states the evaluation work for Chinese mandarin text-to-speech system.The goal of the work is to give a whole assessment to the Chinese mandarintext-to-speech system. Beside the goal we also expect to find the problem exist inthe system and to evaluation the quality of the synthetic speech, and to promote theresearch and the development of Chinese mandarin text-to-speech system.First, the structure of the concatenate-based speech synthesis system is analyzed. Andafter the analyses, assignment of assessment and resolvent is given.Next, an overview of the methods used for the assessment of speech synthesis systemin the world is introduced. Two aspects of the assessment are of linguistics and ofphonetics. The assessment of linguistics aspect aims to the module of text analyzed.And the assessment of phonetics aspect aims to the quality of the synthetic speech bytwo ways - Intelligibility and naturalness. The assessment methods used often aredivided to two sorts - objective method and subjective method.The assessment for linguistics module of Chinese mandarin TTS system needs to begiven notice about characteristics of Chinese mandarin in the sides of segmentation,string of number, polyphone, neutral and retroflex, symbol. The characteristics ofChinese mandarin in the phonetics aspect are also concerned. And the principles ofthe design of the text used in the assessment are given.The assessment methods for phonetics aspect of Chinese mandarin TTS system usedoften are subjective methods. The intelligibility of synthetic speech is assessed by theway of dictation in the two layers - words and sentences. The naturalness of syntheticspeech is assessed by two ways – Mean Opinion Scale (MOS) and Compared Sort(CS). By the first way, synthetic speech is given score which assigned to the level ofthe speech quality by the listeners. By the second way, the listeners compare thequality of some synthetic speech and give sort of them by the quality.The objective assessment method of naturalness for concatenate speech synthesis isresearched. In A Method considering the prosodic parameters of speech, the objectivedistance of patch parameters and duration parameters and intensity parametersbetween the natural speech and the synthetic speech are calculated. For mismatch oftwo speeches in the duration, the DTW (dynamic time warping) algorithm is used toallow approximate matching. In the other Method, objective distance of Mel CepstralCoefficients (MFCC) between the two speeches is calculated. The DTW algorithm isalso use to match the different in the duration. The Mean Opinion Score (MOS)obtained subjectively is compared with the result of the objective assessment.