Ales Prazák

Efficient combination of n-gram language models and recognition grammars in real-time LVCSR decoder

The paper presents a method for incorporation of regular grammars into n-gram language models. Such composite model then benefits from both language modeling formalisms - a grammar yields robust probability estimates for well-defined phrases with fixed structure whereas the n-gram provides better coverage of casual speech. Moreover, the grammar allows adding new words to the phrase pattern while taking advantage of the existing structural (context) information. The proposed method for grammar incorporation allows the use of combined models in our in-house real-time decoder which is designed to work only with standard n-gram language model. The performance of the combined model was tested in the dictation task where a simple grammar was designed for date entries. A statistically significant improvement of WER was achieved.

Multiple application of the MLLT based on clustering supported by phonetic knowledge

This paper intends to show the effect of multiple use of a Maximum Likelihood Linear Transform (MLLT) in continuous speech recognition tasks. The corresponding covariance matrices that were subjected to the transform were clustered using phonetic knowledge. The impact of this approach was tested on the ASRs with PLP- and MFCC-based front-ends and was compared with other known feature decorrelation and dimension reduction techniques.