“CONNECTED DIGIT RECOGNITION IN LOW BIT RATE CODING”

Abstract

This thesis deals with the recognition of digits uttered in continuous manner in noisy coded environment (i.e uttering a telephonic data like phone number). Experiments are carried out in Nepali language, but limited to ten digits (0-9). First acoustic training model with clean data is constructed. Testing for clean data shows 100% recognition. With noisy, coded and noisy coded conditions recognition performance degraded significantly. But with spectral preprocessing method it yields better recognition performance. Three different noises (babble, factory and machine gun) with three different signal to noise ratio (10dB, 15dB and 20dB) are used for noise addition. Babble noise with 10 dB SNR (signal to noise ratio) has lowest recognition rate whereas machine gun noise with 20 dB SNR has highest recognition percentage. GSM (Global System for Mobile Communication) and CELP (Code Book Excited Linear Prediction) are used for coding.