Speech enhancement using DFT and DHT with reduction of cross spectral terms

This paper considered the main limitation of spectral subtraction technique that is, the presence of musical noise in enhanced speech. One of the main reasons for this problem is an incorrect assumption of cross spectral terms involving clean spectrum and noise spectrum as equal to zero. In this paper a new speech enhancement technique based on spectral subtraction is proposed. The principle of proposed method is almost same as simple spectral subtraction with only difference that, in the proposed method the equation of enhanced speech power spectrum not only contains noisy power spectrum and noise power spectrum, as in simple spectral subtraction, but also contains the product term of noisy spectrum and noise spectrum. In proposed method cross spectral terms are estimated using equation contains product of noisy spectrum and noise spectrum. Another objective of this paper is to implement proposed algorithm in such a way that the proposed algorithm execution time should be less and consume less memory. To achieve this, the proposed method is implemented using Fast Hartley Transform (FHT). To evaluate performance of proposed method objective quality measure and subjective listening test are done. Results from these methods show the improvement in performance of proposed method than simple spectral subtraction technique.