A hierarchical two-level analysis structure for use in speech coding and recognition

The authors describe an analysis network that has been successfully applied to redundancy extraction in speech coding and for classification purposes in speech recognition. The network is designed as a two-level hierarchical structure with one building block. This building block performs a DCT (discrete cosine transformation) followed by magnitude and phase-derivative time trajectory computation. A combined short- and long-term analysis is obtained as well as energy compaction to a very small number of characteristic parameters. The parameters reflect both formant and pitch structure of voiced speech signals. The analysis applies to any quasi-stationary signal and is not dependent on any speech production model