Development and Integration of the LDA-Toolkit into the COST249 SpeechDat (II) SIG Reference Recognizer

Bojan Kotnik, Zdravko Kačič, Bogomir Horvat

University of Maribor, Faculty of Electrical Engineering and Computer Science

P27-SE

This paper presents the development of Linear Discriminant Analysis toolkit (LDA-Toolkit) and its integration into widely used COST249 SpeechDat(II) Task Force Reference Recognizer (RefRec). The crucial parts of the LDA, the determination of LDA classes, as well as the influence of the level of dimensionality reduction on automatic speech recognition performance, are discussed. Evaluation of proposed LDA-RefRec procedure is performed using the Slovenian, German, and Spanish SpeechDat (II) databases. HTK (Hidden Markov Model Toolkit) is used in training and recognition processes. Features are computed using Advanced Front End (AFE) feature extraction procedure, proposed by Motorola, France Telecom, and Alcatel (AFE has been also standardized by ETSI organization). Automatic speech recognition results achieved with LDA-RefRec procedure show performance improvement and simultaneously dimensionality reduction when compared to baseline RefRec procedure. Proposed multilingual LDA classes, equal for all the three databases, perform only slightly worse than monolingual LDA classes, constructed and used separately for particular database. The results show benefits of the usage of the proposed LDA-RefRec procedure for evaluation or development of the automatic speech recognition systems based on SpeechDat (II) compliant databases.

Automatic Speech recognition, SpeechDat(II), RefRec Reference Recognizer, Linear Discriminant Analysis, Advanced Front-End

59.pdf