Speech enhancement and recognition of compressed speech signal in noisy reverberant conditions

Abstract

Speech compression, enhancement and recognition in noisy, reverberant conditions is a challenging task. In this paper a new approach to this problem, which is developed in the framework of probabilistic random modeling. speech coding techniques are commonly used in low bit rate analysis and synthesis. Coding algorithms seek to minimize the bit rate in the digital representation of a signal without an objectionable loss of signal quality in the process. Speech enhancement aims to improve speech quality by using various algorithms This paper deals with multistage vector quantization technique used for coding of narrow band speech signals. The parameter used for coding of speech signals are the line spectral frequencies, so as to ensure filter stability after quantization. A new approach incorporates the information about statistical random nature of uncompressed speech signal using LBG algorithm.The code books used for quantization are generated by using Linde, Buzo and Gray(LBG) algorithm. Speech model is characterized by LPC coefficients and parameterized by the coefficients of the reverberation filtersThe results of the multistage vector quantizer are compared with unconstrained vector quantization Technique. The performance of quantization is measured in terms of spectral distortion measured in dB, Computational complexity measured in KFlops and Memory Requirements measured in Floats. From the results it can be proved that multistage vector quantization is having better spectral distortion performance, less computational complexity and memory requirements when compared to unconstrained vector quantization. The proposed approach yields significantly estimating the parameters from the data, better performance in both signal to noise ratio and subjective filter methods. © 2012 Springer-Verlag GmbH Berlin Heidelberg.