Feature selection for robust automatic speech recognition: a temporal offset approach

Abstract

Automatic speech recognition relies on extracting features at fixed intervals. In order to enhance these features with dynamical (delta) components, discrete derivatives are usually computed and added as features. However, derivative operations tend to be susceptible to noise. Our proposed method alleviates this problem by replacing these derivatives with nearby features selected on a per-frequency basis. In particular, we noted that, at low frequency, consecutive samples are highly correlated and more information can be gathered by looking at features farther away in time. We thus propose a strategy to perform this frequency-based selection and evaluate it 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 $$32.10\,\%$$32.10% in accuracy, with most gains in very noisy environments where the traditional delta features have low recognition rates.