Speech Analysis and Processing for Knowledge Discovery Aalborg, Denmark June 4-6, 2008

Time-Varying Cepstral Coefficients

Trond Skogstad, Torbjørn Svendsen

Department of Electronics and Telecommunications, Norwegian University of Science and Technology, Norway

This paper introduces a new set of cepstral features, based on a slightly modified version of the time-varying linear predictive models pioneered by Subba Rao and Liporace. In these models, the non-stationarity of the speech signal is accommodated by expressing the filter coefficients as a weighted combination of known basis functions. By running the parameterized filter coefficients through the recursive link between all-pole models and cepstral coefficients, we obtain a time-varying cepstral representation in analytical form. In a preliminary recognition experiment this representation is shown to give a satisfactory performance. It is argued that the introduced features are well suited for tasks such as detection of landmarks and stationary segments.

Full Paper
Presentation

Bibliographic reference.  Skogstad, Trond / Svendsen, Torbjørn (2008): "Time-varying cepstral coefficients", In SPKD-2008, paper 053.