NGSLT Speech and Speaker Recognition course (Spring 2007) Phonetic Modeling in ASR (Russian speech) - the impact on performance
Valentin A. Smirnov · 2007
In this term paper for NGSLT course “Speech and Speaker Recognition ” we will describe the method for pronunciation modeling, which makes part of automatic speech recognition (ASR) technology for Russian, developed by R’n’D department at Vocative, ltd, where the author works. The method comprises both canonical transcription generation and pronunciation variation modeling. The key idea of the current work is to automate the process of lexicon generation for a given recognition task and to inspect the impact of modeling pronunciation variation on ASR performance. Specific issues concerning pronunciation variation in Russian are discussed and the results on several recognition tasks are presented. 1.