Development of Large Vocabulary Continuous Speech Recognition for Polish

摘要

In this study, the results of acoustic modeling used in a large vocabulary continuous speech recognition system are presented. The acoustic models have been developed with the use of a phonetically controlled large corpus of contemporary spoken Polish. Evaluation experiments showed that relatively good speech recognition results may be obtained with adequate training material, taking into account: (a) the presence of lexical stress; (b) speech styles (a variety of segmental and prosodic structures, various degrees of spontaneity of speech (spontaneous vs. read speech), pronunciation variants and dialects); (c) the influence of the sound level and background noises. The present large vocabulary continuous speech recognition evaluation results were obtained with Sclite assessment software. Moreover, the article delivers information about the speech corpus structure and contents and also a brief outline of the design and architecture of the automatic speech recognition system.