This paper describes a new segmentation algorithm for the "Neural" Phonetic Typewriter developed at Helsinki University of Technology. The algorithm is based on use of homogeneity and length of speech segments as their reliability parameters. The maximum reliability is sought b}r merging short segments, which increases individual segment lengths and possibly decreases their homogeneity. The merging leads to formation of tree structure. The final stage of the algorithm is the search for an optimal path through it. Before segmentation, vector-quantized representation of speech is produced with a codebook created by Learning Vector Quantization. Each codebook vector is associated with another vector describing probabilities of the phonemic classes. The sum of these phoneme probability vectors within a segment determines its classification and homogeneity.
Bibliographic reference. Laaksonen, Jorma T. (1991): "A new reliability-based phoneme segmentation method for the "neural" phonetic typewriter", In EUROSPEECH-1991, 97-100.