Uniform Concatenative Excitation Model for Synthesising Speech without Voiced/Unvoiced Classification

摘要

In general, speech synthesis using the source-filter model of speech production requires the classification of speech into two classes (voiced and unvoiced) which is prone to errors. For voiced speech, the input of the synthesis filter is an approximately periodic excitation, whereas it is a noise signal for unvoiced. This paper proposes an excitation model which can be used to synthesise both voiced and unvoiced speech, thus overcoming the problem of degradation in speech quality caused by those classification errors. Basically this model consists of representing two contiguous segments of the residual signal pitchsynchronously. The first segment is represented by the original residual in a fraction of the period around the pitch-mark (obtained using an epoch detector), in order to capture the most important aspects of the residual during voiced speech. Instead, the remaining part of the period is modelled by a set of parameters of the amplitude envelope of the residual waveform and its energy. The technique for synthesising the excitation combines these shaping parameters with a novel method for regeneration of the residual waveform and a method to mix a periodic signal with noise based on the Harmonic plus Noise model. Besides producing high-quality speech, this technique is computationally fast.