Implementation of high-quality text-to-speech synthesis for limited domain applications

Abstract

Though current TTS systems can produce unrestricted, intelligible speech using a limited database of prerecorded sounds, the quality of speech is highly artificial. By using larger speech segments, such as a combination of words and diphones, the number of segment boundaries are reduced drastically, resulting in high-quality TTS that has a humanlike natural voice. Such a system can be useful for limited domain applications such as driving directions, weather advisories, etc. Experiments were conducted to generate natural sounding speech from words and diphones by modifying the pitch and duration of the units to match the prosody characteristics. The Festival TTS system, developed at the CSTR at the University of Edinburgh, was used to aid in the implementation of the experiments. The method for synthesis is pitch-synchronous residual LP. This method allows for producing smooth transitions between segments. A database of residuals and parameters for segments (words and diphones) is prepared. During synthesis, the segments are concatenated and the desired prosody is matched using time-domain modifications of pitch and duration. Using natural prosody, waveforms were synthesized, resulting in close to natural sounding speech. This demonstrates synthesized speech can sound identical to the original speaker, given the correct prosody.