Digital Signal Processing

Yogesh S. Angal ¹, Amol D. Rahulkar ², Raghunath S. Holambe ³, Rajan H. Chile ³

Affiliations
1 Department of Instrumentation Engineering, Padmashree Dr. D.Y. Patil Institute of Engineering and Technology, Pimpri, Pune-18, Maharashtra State (M.S.), IN
2 S.G.G.S. Institute of Engineering and Technology, Nanded, IN
3 Department of Instrumentation Engineering, S.G.G.S. Institute of Engineering and Technology, Nanded-431607, Maharashtra State (M.S.), IN

Abstract

This paper presents a new technique to extract the speech features in order to improve the recognition accuracy in various types of noisy environments. Most of the speech recognition systems are suffered from high computational complexity. In this paper, a new class of triplet half band wavelet packets (THWP) has been designed based on the generalized half band polynomial. These packets are used in speech recognition system to derive the effective and efficient speech features. The proposed THWP satisfies perfect reconstruction (PR) and provides linear phase, regularity, better frequency-selectivity and near orthogonality. These properties are exploited to approximate desirable speech features significantly. The proposed technique computes features using energy, mean and variance of each sub-band of THWP. This gives low dimensional feature vectors for speech recognition purpose. The performance of the proposed algorithm has been evaluated on Texas Instruments-46 (TI-46) speech database in various noisy environments. The performance of the proposed technique is better than existing popular speech recognition algorithms.

Keywords

Filter Bank, Half Band Filters, Feature Extraction, Wavelet Transform, THWP, Speech Recognition.

Full Text

     

Yogesh S. Angal ¹, Mangesh S. Deshpande ², Raghunath S. Holambe ³, Rajan H. Chile ³

Affiliations
1 Department of Instrumentation Engineering, Padmashree Dr. D.Y. Patil Institute of Engineering and Technology, Pimpri, Pune-18, Maharashtra State (M.S.), IN
2 S.G.G.S. Institute of Engineering and Technolgy, Nanded-431607, Maharashtra State (M.S.), IN
3 Department of Instrumentation Engineering, S.G.G.S. Institute of Engineering and Technology, Nanded-431607, Maharashtra State (M.S.), IN

Abstract

This paper presents a new feature extraction technique based on a Gammatone Filter Bank (GTFB) for speech recognition using Radon Transform (RT) and Discrete Cosine Transform (DCT). In the proposed scheme speech specific features have been extracted by applying image processing technique to the patterns available from speech signal by applying Gammatone Filter Bank. Radon projections for twenty six orientations are captured. The acoustic characteristics of the Gammatone Filter Bank applied to the speech signal. DCT applied on Radon projections yields low dimensional feature vectors. The technique is computationally efficient and robust to session variations and insensitive to additive noise. The performance of the proposed algorithm is evaluated in presence of additive white Gaussian noise from (30dB to -5dB SNR) on Texas Instruments-46 (TI-46) speech database. The proposed algorithm improves the performance of the speech recognition system in noisy environment compared to the existing popular algorithms like Mel frequency Cepstral Coefficient (MFCC), Linear Predictive Cepstral Coefficients (LPCC), Perceptual Linear Prediction (PLP).

Keywords

Speech Recognition, Gammatone Filters, Feature Extraction, Radon Transform, Discrete Cosine Transform.

Full Text

     

Yogesh S. Angal ¹, Pawan K. Ajmera ², Raghunath S. Holambe ³, Rajan H. Chile ⁴

Affiliations
1 Department of Instrumentation Engineering, Padmashree Dr.D.Y. Patil Institute of Engineering and Technology, Pimpri, Pune-18, Maharashtra, IN
2 S.G.G.S. Institute of Engineering and Technolgy, Nanded. Maharashtra State (M.S.), IN
3 Department of Instrumentation Engineering at S.G.G.S. Institute of Engineering and Technology, Nanded Maharashtra State (M.S.), IN
4 Department of Instrumentation Engineering at S.G.G.S. Institute of Engineering and Technology, Nanded Maharashtra State (M.S.) 431 607, IN

Abstract

This paper presents a new feature extraction technique for speech recognition using Radon Transform (RT) and Discrete Cosine Transform (DCT). A spectrogram is a time varying spectrum(forming an image) that shows how the spectral density of a signal varies with time. In the proposed scheme speech specific features have been extracted by applying image processing technique to the patterns available in the spectrogram. Radon transform has been used to derive the effective acoustic features from speech spectrogram. The proposed technique computes radon projections for nine orientations and captures the acoustic characteristics of the speech spectrogram. DCT applied on Radon projections yields low dimensional feature vectors. The technique is computationally efficient, speaker-independent, robust to session variations and insensitive to additive noise. Radon projections for seven orientations capture the acoustic characteristics of the spectrogram. The performance of the proposed algorithm has been evaluated in presence of additive white Gaussian noise from (30dB to -5dB SNR) on Texas Instruments-46(TI-46) speech database. The performance of the proposed technique in noisy environment is much better than existing popular algorithms.

Keywords

Speech Recognition, Spectrogram, Feature Extraction, Radon Transform, Discrete Cosine Transform.

Full Text