Digit Identification from Speech using Short-Time Domain Features

Digit identification from speech has numerous applications such as automatic voice dialing, audio based pin entry, reservation in airlines, banking etc. In order to recognize digits, analysis was carried out with the help of audio containing digits from 0-9, recorded in normal environmental conditions. For this, a database consisting of 100 samples of each digit are acquired with the help of native subjects in Tamil language, at sampling frequency 44 kHz. Pre-processing techniques were done for silence and noise removal. Primary features: short-time energy and short-time zero crossing count were extracted. Further the statistical parameters from these features: mean, variance, 25%, 50%, 75% and 100% quartile values were used to recognize the digits. Conventional machine learning methods: K nearest neighbour, support vector machines using linear and radial basis functions, Naïve Bayes, decision tree, Adaboost and random forest classifiers are used to identify digits from audio input. The work also showed that pre-processing, by silence removal, could bring significant improvements in classifier performance, for the digit dataset under analysis.