MELP vocoder using lifting wavelet transform

A MELP (mixed excitation linear prediction) vocoder produces much improved voice quality compared with the traditional LPC vocoder. An LPC synthesis filter which mimics the vocal tract characteristics is excited with a mixture of pulse and noise instead of using either of them. This paper describes an implementation of the MELP vocoder which uses the lifting wavelet transform in place of the band-pass filter bank required for subband division in the synthesis and analysis of the MELP vocoder. A new method to generate an appropriate glottal waveform is also described. In addition, three kinds of fluctuations observed in the steady parts of voiced speech are incorporated to enhance the naturalness of synthesized speech. This MELP vocoder was realized on a DSP evaluation module to confirm the feasibility of real-time implementation.