DIPHONE SPEECH SYNTHESIS

Text-to-speech synthesis requires two steps: linguistic processing to convert text into phonemes and intonation parameters and simulation of speech production (to generate the speech waveform). We review different methods for the second task, emphasizing the advantages and disadvantages of the linear predictive (LPC) diphone approach. Diphones require more memory to represent all possible spectral transitions between pairs of phonemes, but they directly capture many of the coarticulation effects that must otherwise be modeled in phonemic synthesis. Relatively simple interpolation is allowed due to the similarity of spectra at diphone boundaries.