ESSV Konferenz Elektronische Sprachsignalverarbeitung

Title: Training of HMMs for Pronunciation Error Detection – Crosslingual Bootstrapping vs. Flatstart Training

Authors: Michael Beilig, Diane Hirschfeld, Oliver Jokisch, Uwe Koloska

Abstract:

This paper presents an evaluation study of hidden markov model (HMM) training as part of the ongoing EURONOUNCE project that intends an Intelligent Language Tutoring System (ILTS). The Core functionality of the system is a computer assisted pronunciation trainer (CAPT), integrating audio–visual feedback based on a pronunciation error detection on the phone–level. In a first step the application will concentrate on Slavonic languages and German. Most of current approaches solely use hidden markov models (HMMs) of the language to learn (target language models). One characteristic of our approach is to additionally use source language as well as intermediate acoustical models to detect and rate pronunciation errors. The other specific is that our methods to achieve this goal are directly based on knowledge of linguistic experts and experiences of pronunciation training. So the extended acoustic model inventory is used by an expert to formulate pronunciation error hypotheses for each of the training utterances strongly related to the source–target language (L1–L2) pair just as to specific learning levels. As an initial study a cross language bootstrapping approach for German as sourceand Polish as target–language was implemented to train a native Polish monophone model set. By this way, the authors approximate the break–even amount of data, at which the cross language bootstrapping can outperform the flatstart training procedure.


Year: 2009
In session: Sprache und Didaktik
Pages: 372 to 379