Sixth European Conference on Speech Communication and Technology
The popularity of Line Spectra Pairs (LSP) in speech processing has been supported recently by theoretical studies of their statistical properties. It has been shown that LSP frequencies are uncorrelated, and have a diagonal sensitivity matrix with respect to spectral distortion. Therefore, LSP is suitable for Vector Quantization (VQ) schemes with simple weighted Euclidean distance measures. This was further supported by an analysis of the LSP probability density function, shown to be appropriate for distance-based recognition frameworks. This paper reports our study on developing improved methods for using LSP in VQ based speaker recognition. We used Linear Discriminant Analysis (LDA) to explore the speaker-discrimination statistics of LSP, and to transform them into a speaker-discriminative space. Further enhancements include the use of special VQ distance measures such as F-ratio weighting and Inverse Harmonic Measure (IHM). Performance evaluation experiments were conducted on very short speech sessions, using a database of 32 male speakers, taken from the TIMIT and NTIMIT. Identification results of 100% for clean speech and 70% for telephone speech were achieved.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Zilca, R. D. / Bistritz, Y. (1999): "Text independent speaker identification using LSP codebook speaker models and linear discriminant functions", In EUROSPEECH'99, 799-802.