Iterative parameter optimization techniques for parallel-formant encoding of speech

One of the major difficulties with the extraction of synthesiser control parameters from natural speech arises from formant interactions in the synthesised speech, both within a single frame, and energy carry-over between frames. In this paper, iterative parameter optimization techniques have been used to investigate these problems. To reduce the computational load inherent in such an approach, the optimization algorithms were developed on a PC with a plug-in board containing a real-time DSP-based synthesiser. Locally-based linear predictive coding (LPC) spectral error measures have been developed, and used to design algorithms to optimize both frequency and amplitude synthesiser parameters. Examples of typical error surfaces are studied and results of parameter optimization given