Automatic speech recognition in adverse acoustic conditions

Más información sobre el libro

Most of today`s speech recognition systems are based on a cepstral analysis scheme for representing short-term speech features. Furhtermore, the statistical approach of modeling speech units by Hidden Markov Models (HMMs) and the application of the Viterbi algorithm are state-of-the-art for the pattern recognition in this field. These analysis and modeling techniques are presented as an introduction to the problem of recognizing speech in adverse acoustic conditions. Analyzing the speech input in application scenarios of recognition systems the speech signal is affected by the acoustic environment. An overview about the combination of possible distortion effects is given. A simulation tool is presented that has been developed to allow the generation of distorted speech data for a lot of different acoustic environments. A new database is described that has been created with this tool. It is publicly available for comparative investigations on the recognition of distorted speech data.