Segment vocoder based on reconstruction with natural segments

A segment vocoder based on a codebook of natural (non-warped) speech segments has been constructed. In contrast to other segment vocoders, in which segments of various lengths are generated by time-warping, the new vocoder uses only non-warped segments, whose lengths are determined during the segmentation process. The vocoder also incorporates an improved segment distance measure and compensates for input block-boundary effects. To generate codebook templates, the authors use an iterative improvement method to segment training speech. The vocoder achieves 85% DRT (diagnostic rhyme test) intelligibility with non-warped templates compared to 83% DRT with warped templates. A baseline segment vocoder that included none of the improvements was only able to achieve 77% DRT.<<ETX>>