engleski

INTRODUCTION TO SPEECH RECOGNITION, Exercise in ASR using HTK

These course notes are intended to cover the practical part of the course “Introduction to speech recognition”. The course covers the basics about hidden Markov models (HMM), how to build a new acoustic model for English language using HMMs and how to test the quality of this model (recognition accuracy). Model testing will be performed on utterances with limited vocabulary and strictly defined word-to-word transitions. Utterances will be recorded during these exercises by each student individually. Finally, quality and complexity of the acoustic model will be compared with another model, which will be constructed from all the utterances of all students. Furthermore, students will learn how to work with the Hidden Markov Model Toolkit (HTK) which is a widespread tool for building and testing of acoustic and language models. It provides an opportunity to build models from scratch and more importantly, step by step, thus gaining insight into the structure of a typical ASR system. The main goal of these exercises is to teach the students to research and explore the ASR world and gain hands-on experience using their own speech examples.

Automatic speech recognition; ASR; HTK

Course on „Automatic speech recognition“ (ASR) was prepared and given as a one of 12 courses given on the summer school entitled: “Interdisciplinary Summer School, Workshop and Round Table in Computational Linguistics, Cognitive and Information Science”, Zadar, 2010. The course was comprised of 5 hours of lectures, giving theoretical background of ASR, and 5 hours of exercises (covered by these course notes). All exercises were based on individual student work, using Hidden Markov Model Toolkit (HTK). The course covered main aspects of ASR systems, such as feature vector extraction, building of acoustic models, actual recognition as well as recognition accuracy evaluation.