Academic Journals Database
Disseminating quality controlled scientific knowledge

Pattern Normalization/Template Optimization in Order To Optimize Speech Recognition Process

Author(s): Mutcha Srinivasa Rao

Journal: International Journal of Scientific Research and Reviews
ISSN 2279-0543

Volume: 1;
Issue: 2;
Start page: 69;
Date: 2012;

Keywords: Pattern Normalization | Template Optimization | Speech Recognition | ASR | Speech Optimization | Dynamic Time Warping | Time Normalization.

One of the main problems in speech recognition systems is the preparation of reliable referencetemplates for the set of words to be recognized. The accuracy of the speech recognition systems greatlyrelies on the quality of the prepared reference templates. The normal procedure in selecting thereference templates is to select one example then test its recognition rate. If the recognition rate is highthen this reference is kept, otherwise another template has to be selected.A common way to improve the recognition performance is to use several templates for each word. Thisprocedure is computationally inefficient because it increases the number of templates. Vectorquantization (VQ) is another solution to prepare reliable templates for the DTW-based speechrecognition systems. However, it requires many training examples to prepare a reliable codebook. Inorder to keep up the computational efficiency feasible, a simple method is to use single referencetemplates per word. Even if single templates per word are chosen, there are two disadvantages: A single sample template per utterance cannot account for the variability of the speech signal.Even the same user cannot speak the same word exactly the same each time. There is no way to indicate the quality of information content contained in the sample template.The solution to this problem is to make use of an Average Template Method that has the followingfeatures: Many samples are taken initially to account for variability in the speech signal for each word. It extracts the reference template from a set of examples rather than one example. The effect of a bad sample may be mitigated out of the good samples. No additional computational cost incurred during the overall recognition Process.
Affiliate Program     

Tango Rapperswil
Tango Rapperswil