|
In this thesis, a CELP-based text-to-speech conversion system
is presented. We take 1410 Mandarin Chinese monosyllables as
the basic synthetic units in this system. The Code Excited
Linear Prediction (CELP) algorithm is applied to our speech
synthesizer for high compression rate and good speech quality.
In order to improve the naturalness of the synthetic speech, a
method for prosodic modification is proposed to replace the
traditional rule-based approach for pronunciation. At first, a
total of 12 representative pitch contour patterns are defined
for the behavior of four lexical tones and a neutral tone in
Mandarin Chinese. By the observation, it appears that the
acoustic properties of a syllable may be affected by the
different concatenation condition in a sentence. Consequently,
a Bayesian network is employed to model the relation between
fluctuation of pitch contour and linguistic features. This
network is trained by a set of sentence utterance and provides
appropriate prosodic information for adjusting the synthetic
speech in the synthesis process. The synthetic speech has been
tested on 20 subjects. The results indicated that the average
correct rate is 96.65% for intelligibility, and the ratio for
the mean opinion score above "fair" level is 84.31% for
naturalness.
|