First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Line Spectrum Pair Frequency - Based Distance Measures for Speech Recognition

Fikret S. Gurgen, Shigeki Sagayama, Sadaoki Furui

NTT Human Interface Laboratories, Speech and Acoustics Lab., Tokyo, Japan

In the present study, the performance of the line spectrum pair (LSP) frequencies representation for speech recognition is investigated. Various distance measures such as Euclidean, inverse variance weighted Euclidean, and Mel-scale-like weighted distance measures based on the LSP frequencies are used for speaker-independent isolated word recognition experiments with a Dynamic Time Warping (DTW) system. Transitional LSP frequency parameters defined by regression coefficients of LSP frequencies are also introduced. The transitional and the instantaneous parameters and distances are linearly combined for better recognition performance. The cepstral distance measures, and transitional and instantaneous cepstral parameters and distances are used for the comparison of the performances. The linear combination of the instantaneous and the transitional parameters for LSP representation is found to be the best among the all distances used in the experiments.

Full Paper

Bibliographic reference.  Gurgen, Fikret S. / Sagayama, Shigeki / Furui, Sadaoki (1990): "Line spectrum pair frequency - based distance measures for speech recognition", In ICSLP-1990, 521-524.