Branislav Gerazov and Zoran Ivanovski


Automatic Speech Recognition Systems of today are intensely deployed in real world application scenarios which are often characterized by suboptimal operating conditions. Thus their noise robustness has become a crucial parameter when assessing ASR in-field performance. The paper examines the noise robustness of traditional ASR feature sets as applied to a Voice Dialing Application built for Macedonian. The analysis focused on the following features: Linear Prediction Reflection Coefficients, Mel-Cepstral Cepstral Coefficients and Perceptual Linear Prediction Coefficients. The ASR system was trained with clean data, and in the evaluation phase the noise level in the test data was varied by adding white and babble noise. Results have been plotted for each feature type across varying SNR conditions.


ASR, features, noise robust, noise, SNR