Cascaded Structure of Wiener Filter with FBS based Spectral Splitting and Dynamic Range Compression For Listeners with Sensorineural Hearing Loss

This paper presents a cascaded structure of wiener filter followed by spectral splitting using Filter Bank Summation (FBS) method and dynamic range compression with constant compression factor for binaural dichotic presentation. This helps to compensate the effect of spectral masking, reduced dynamic range for moderate sensorineural hearing loss and also improves speech perception in the adverse listening conditions by removing unwanted noise. Wiener filter is used as noise reduction algorithm. Using auditory critical bandwidths, the signal is separated into eighteen frequency bands, varying from 0-5 KHz. Spectral splitting is performed by combing odd and even numbered filter bands and presented to the left and right ears respectively that help in reducing the effect of frequency masking. Amplitude compression using fixed compression factor (0.4, 0.6 and 0.8) in each band was carried out that helps in compensating the reduced dynamic range of the hearing impaired people. Listening tests using Mean Opinion Score (MOS) on normal-hearing subjects were conducted to test the perceived quality of the processed speech. VCV syllable /aba/ and sentences were used as the test material. The results of the listening tests showed that the MOS scores for the processed speech (4.31, 4.06, 3.68, 3.5, 3.067 and 2.612) were higher as compared with those for unprocessed speech (4.48, 1.2, 1.13, 1.06, 0.8 and 0.45) for SNR values of ∞, +6, +3, 0, -3 and -6 dB respectively, indicating the improvement in the perceived quality for different SNR conditions for the compression factor of 0.6. To estimate the intelligibility of the perceived speech, listening test was also carried out on hearing impaired (moderate SNHL) subjects in the presence of background noise using Modified Rhyme Test (MRT).The test material consists 50 sets of monosyllabic words of consonant-vowel-consonant (CVC) form with six words in each set. Each subject is responded for total of 1800 presentations (300 words x 6 different SNR conditions). Results of the listening tests (using MRT) showed maximum improvement in speech recognition scores (23.49 - 27.29%) compared to unprocessed speech, indicating enhancement in intelligibility of the speech at lower SNR values.