A New Spectral Subtraction Method for Speech Enhancement using Adaptive Noise Estimation

Speech enhancement is the basic requirement for speech signal based real life applications such as automatic speech recognition, speaker identification system. Improving the quality and/or intelligibility of degraded/noisy speech signal has been a topic of interest for the last many decades. Main objective of speech enhancement is to improve the perceptual aspects of the speech such as overall quality, intelligibility and degree of listener fatigue. The additive stationary/non-stationary background noise degrades the quality of transmitted/recorded speech. Existing methods like spectral subtraction, Cepstral mean subtraction, blind equalization etc., are being used for speech enhancement. Basic spectral subtraction method is the oldest one for reducing the additive background noise in speech signal. In this approach noise spectrum is estimated at silence region at the beginning of the speech and assumed that noise will be stationary throughout the speech. Actually, in real life noise spectrum shows non-stationary behavior during the speech. The proposed algorithm in this paper continuously updates the noise spectrum on the basis of short term energy (STE). Due to the adaptive estimation capability of noise spectrum, it works well for stationary as well as non-stationary noise. Noise spectrum is continuously updated throughout the speech by comparing short term energy value of the current frame to the average short term energy values of all the preceding frames. Many experiments are performed to verify the performance of the proposed method. Under different noise conditions, the performance of the developed algorithm has been compared with existing spectral subtraction methods. Simulations show that the proposed spectral subtraction using adaptive noise estimation method produce cleaned speech signal with better SNR values irrespective of the source of noise.