Temporal Feature Selection for Noisy Speech Recognition

Automatic speech recognition systems rely on feature extraction techniques to improve their performance. Static features obtained from each frame are usually enhanced with dynamical components using derivative operations (delta features). However, the susceptibility to noise of the derivative impacts on the accuracy of the recognition in noisy environments. We propose an alternative to the delta features by selecting coefficients from adjacent frames based on frequency. We noticed that consecutive samples were highly correlated at low frequency and more representative dynamics could be incorporated by looking farther away in time. The strategy we developed to perform this frequency-based selection was evaluated on the Aurora 2 continuous-digits and connected-digits tasks using MFCC, PLPCC and LPCC standard features. The results of our experimentations show that our strategy achieved an average relative improvement of \(32.10\%\) in accuracy, with most gains in very noisy environments where the traditional delta features have low recognition rates.