International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169 Volume: 11 Issue: 6 DOI: https://doi.org/10.17762/ijritcc.v11i6.7059 Article Received: 18 March 2023 Revised: 26 April 2023 Accepted: 20 May 2023 ___________________________________________________________________________________________________________________ 71 IJRITCC | June 2023, Available @ http://www.ijritcc.org Machine Learning Based Dynamic Band Selection for Splitting Auditory Signals to Reduce Inner Ear Hearing Losses Sudhir Narsing Divekar 1 , Dr. Manoj Kumar Nigam 2 1 Department of Electronics Engineering, MATS University, Raipur (C.G.) sndivekar.pp@gmail.com 2 Professor, Department of Electrical & Electronics Engineering, MATS University, Raipur (C.G.) drmanojk@matsuniversity.ac.in Abstract—Quality of hearing has been severely impacted due to signal losses occurs in the human inner ear explicitly in the region of cochlea. Loudness recruitment, degraded frequency selectivity and auditory masking are the major outward effects of inner ear hearing losses. Splitting auditory signals into frequency bands and presenting dichotically to both ears became a comprehensive solution to reduce inner ear hearing losses. However, these methods divide input signal into the fix number of frequency bands, this limits their applicability where signals have large variations in their spectral characteristics. To address this challenge, we have proposed machine learning based intelligent band selection algorithm to split auditory signals dynamically. Proposed algorithm analyze input speech signal based on spectral characteristics to determine the optimum number of bands required to effectively present major acoustic cues of the signal. Further, dynamic splitting algorithm efficiently divides signal for dichotic presentation. Proposed method has been examined on large number of subjects from different age groups and gender having cochlear hearing impairment. Qualitative and quantitative assessment shown significant improvement in the recognition score with substantial reduction in the response time. Index Terms— Inner ear hearing loss, Intelligent band selection, Spectral splitting, Cochlear hearing impairment, Machine learning algorithm. I. INTRODUCTION UMAN ear mainly comprises of outer ear, middle ear and inner ear. Hearing losses associated with outer ear and middle ear are due to loud sound exposure, contagious infections, or due to the head injury. Most often these problems are resolved using proper medication or in case of severity they can be treated surgically [1]. However, hearing losses associated with inner ear are difficult to address using prevailing medical treatments. Inner ear losses are resulted from damage to inner or outer hair cells, hardening of organ of corti or aging can also be a substantive cause for signal losses as auditory functions degenerates more hastily in elderly. Loudness recruitment, degraded frequency selectivity and auditory masking are the major outward effects of cochlear hearing losses. Unfortunately, there is no comprehensive technique available to comprehensively address hearing losses occurs in the human inner ears [2]. To reduce hearing losses in the inner ears explicitly occurring in the region of cochlea, various band splitting methods has been proposed in the literature. These methods split auditory signals to form different bands based on their frequency contents. To split auditory signals, either bank of filters or transforms such as wavelets has popularly been used in the literature. Simultaneously presenting alternate split bands in even-odd manner to both ears significantly minimized the frequency overlapping and helped in improving the audibility in hearing impaired [3]. Though, such methods are efficient when spectral cues of signal are flat and consistent, but they underperform in real time scenario where interference of the background noise introduces unwanted variations and fluctuations in the spectral information of signal. To address this issue, we proposed machine learning based intelligent band selection algorithm to split auditory signals. Proposed algorithm analyze input speech signal based on spectral characteristics such as place and manner of articulation, relations between formant frequencies, voice and unvoiced distinction, and vowel differentiation. Based on this spectral analysis, algorithm determines optimum number of bands required to effectively present major acoustic cues of the signal. Algorithm also fixes the frequency range of each band dynamically. Proposed method has been examined on large number of subjects from different age groups and gender having cochlear hearing impairment. Qualitative and quantitative assessment shown significant improvement in the recognition score with substantial reduction in the response time when compared with state-of-the-art methods. The rest of the paper is organized as follows: Section II discusses existing band splitting methods and their band H