Diphone speech synthesis based on a pitch-adaptive short-time fourier transform
Stephen C. Glinski · 1981
The purpose of this work is to investigate a new method of speech synthesis from phonetic specifications. The investigation includes the design, computer simulation, and subjective evaluation of a speech analysis-synthesis system. The method is new in the sense that it utilizes two novel analytical techniques: (1) discrete pitch-adaptive short-time Fourier analysis, and (2) diphone representation of real speech. The pitch-adaptive transformation is implemented via a sliding rectangular window whose edges are located at zero crossings of the speech signal, and whose length is one pitch period for voiced regions and constant for fricative regions. This approach is shown to result in a more accurate spectral representation and to offer possibilities for data compression. Algorithms are developed for dynamic pitch, intensity, and time axis warping of the signal during synthesis. Using the adaptive transform, the author's voice is analyzed to produce several diphone templates. These templates are concatenated and smoothed to form synthetic English speech. Results indicate that by using aforementioned techniques, it is possible to produce very intelligible synthetic speech which retains, to a limited extent, the voice quality of the original speaker.