Improved analysis/synthesis methods for the multiband excitation coder

The multiband excitation coder (MBE) has been shown to be a suitable speech coding scheme for bit rates ranging from 2400 bps to 8000 bps. However the speech it produces is sometimes rough with a reverberating quality and its performance seems to be speaker dependent. This somewhat reduced quality (specially at low bit rates) is the main drawback of the MBE coder. According to our results loss of quality is mostly due to the following reasons: mistakes of the pitch tracker, wrong voiced/unvoiced decisions, improper spectra for the excitation's noisy part and lack of temporal resolution at the synthesizer. Effects resulting from quantization needs such as roughness of the spectral envelope are also important. In this contribution solutions to all those problems are proposed. They considerably improve final quality over standard MBE coders. This final quality is almost completely speaker independent.<>