ROBUST ISOLATED DIGIT RECOGNITION SYSTEM USING SPECTRAL SUBTRACTION BASED ON MINIMUM STATISTICS

• Main Author: (NIM : 232 07 007); Pembimbing : Dr. Suhartono Tjondronegoro, FITRILINA
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/15711
• Institution: Institut Teknologi Bandung

The need for robust speech recognition system can not be avoided. Noise can cause loss of quality and intelligibility of the speech signal, eventually degrade the performance of speech recognition system. Therefore, in this study developed an isolated Indonesian digit recognition system equipped with speech enhancement techniques. In this system, modeling of speech using HMM and speech enhancement techniques using spectral subtraction with noise estimator based on minimum statistics. Minimum statistics method estimate the noise power using minimum values of a smoothed periodogram of the noisy speech signal. The basic idea of spectral subtraction is to obtain an estimate of the speech spectral by subtracting the noise estimation from noisy speech signal. In this research, each HMM has 13 states and each state modeled by Gaussian Mixture Model (GMM) with 5 gaussian. Testing the system on the clean environment (matched trainingtesting) achieve recognition accuracy of 89.53%. Testing is also done on the environmental condition of AWGN, the sound of hair dryer and the sound of car. Tests conducted at 20 dB, 15 dB, 10 dB, 5 dB, and 0 dB SNRs. The lower of SNR, the greater of reduction in system performance. The decrease in system performance due to AWGN, hairdryer sound noise, and car sound noise at SNR 0 <br /> <br /> <br /> dB are 69.85%, 65.97% and 63.69%, respectively. Enhanced speech signal using spectral subtraction with noise estimator based on minimum statistics can improve recognition accuracy in AWGN noise by 21.37 %, hairdryer sound noise by 20.11%, and car sound noise by 29.48%. Although the performance of system with single spectral subtraction module is smaller than the performance in <br /> <br /> <br /> matched training-testing, but the use of more than one spectral subtraction module does not provide a significant increase in accuracy, so the use of one module would be more effective.