Efficient Speech Denoising Techniques for Adverse Acoustic Conditions

Nowadays, the real time noise suppression systems struggle by introducing temporal lag and perceptual artifacts in the speech assistive devices. This paper proposes an real time auditory-aware noise spectral tracking methodology which merges both the critical bandwidth analysis and the SNR refinement architecture. This proposed method uses decision-directed recursive technique with a Joint Maximum a Posteriori (JMAP) correction which minimizes the frame delay by incorporating frequency-dependent masking threshold derived from Bark-scale critical bands. It applies a noise-to-masking ratio adaptive weight in the noise estimator to modulate the noise spectral suppression depending on the level of the a posteriori SNR values. Results using TIMIT speech database degraded by stationary white noise to highly non stationary babble noise including factory, street, and Himalayan Snow fall noise interference at various noise levels (0 - 20 dB SNR) demonstrated that the developed noise estimator outperformed in terms of PESQ (up to 3.65), STOI (up to 0.985), and segmental SNR against the state of the art noise estimation techniques available till now.