Acoustic model optimisation for a call routing system

Abstract

The paper presents work aimed at optimising acoustic models for the AutoSecretary call routing system. To develop the optimised acoustic models: (1) an appropriate phone set was selected and used to create a pronunciation dictionary, (2) various cepstral normalization techniques were investigated, (3) three South African corpora and multiple training data combinations were used to train the acoustic models, and, (4) model-space transformations were applied. Using an independent testing corpus, which contained proper names and South African language names, a named-language recognition accuracy of 95.11 % and proper name recognition accuracy of 93.31% were obtained.