Novel Environmentally-Aware Speech Enhancement for Improved Speech Intelligibility in Noise for Cochlear Implants

Lead Research Organisation: University College London
Department Name: Electronic and Electrical Engineering

Abstract

Hearing loss costs the UK over £30 billion per year. With most cases of sensorineural hearing loss, there is damage to the sensitive hair cells inside the inner ear that translate sound waves into electrical nerve signals that then travel along the auditory nerve to the brain. By fitting a cochlear implant, hearing can be restored to a certain extent by bypassing the damaged hair cells and stimulating the auditory nerve directly with electrical impulses. However, even with the most sophisticated current cochlear implant technology, most cochlear implant users still find it difficult to understand or follow conversation in a noisy background. One reason for this is because the speech enhancement algorithms deployed in cochlear implants have been created to accommodate all acoustic environments even though real-world auditory environments exhibit a wide variety of temporal and spectral characteristics that require a more adaptable approach to speech enhancement.
This PhD project aims to make use of machine learning to develop an adaptive, compact and robust speech enhancement front-end processor for cochlear implants. The system will have the ability to seamlessly and promptly change its speech enhancement approach depending on the listening environment surrounding the user. The technology will improve the speech intelligibility and quality in noisy environments.

Publications

10 25 50
publication icon
Abdullah S (2020) Comparison of Auditory-Inspired Models Using Machine-Learning for Noise Classification in International Journal of Simulation Systems Science & Technology

publication icon
Abdullah S (2020) Acoustic Noise Classification Using Selective Discrete Wavelet Transform-Based Mel-Frequency Cepstral Coefficient in International Journal of Simulation Systems Science & Technology

publication icon
Zamani M (2020) Dictionary Construction for Accurate and Low-Cost Subspace Learning in Unsupervised Spike Sorting in International Journal of Simulation Systems Science & Technology

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512400/1 01/10/2017 31/03/2022
1946575 Studentship EP/R512400/1 20/11/2017 20/02/2022 Salinna Binti Abdullah
 
Description The work funded through this award investigates methods to improve speech intelligibility and quality for hearing prosthesis users through speech enhancement. Cochlear implant and hearing aid users often suffer from poor speech intelligibility and quality in noisy environments. Speech enhancement, which is the task of separating speech from nonspeech noise, can be used to improve hearing in challenging listening conditions for hearing prosthesis users. A simple speech enhancement method is spectral subtraction, where an estimate of the noise is obtained and subtracted from the noisy speech to give the denoised speech. However, the performance of speech enhancement methods that rely on accurate estimations of the speech or noise components degrades considerably when dealing with non-stationary noise and noisy speech at low SNRs since it is difficult to estimate the speech and noise properties in these conditions accurately. Due to this reason, speech enhancement algorithms utilizing supervised deep-learning approaches have garnered much research attention since they have demonstrated exceptional denoising capability even in the challenging noise conditions mentioned. In the work funded through this award, a deep neural network-based speech enhancement utilizing a quantized correlation mask as its training targets is proposed. The proposed method resulted in improvements across all the objective speech intelligibility and quality scores employed for evaluation when compared to the unprocessed noisy speech as well as the enhanced speech outputs from other methods used for comparisons. The proposed method was able to produce better performance despite efforts implemented to increase its compactness and decrease its computational complexity as the proposed method also demonstrated faster training and inference time, and reduced memory requirement. Besides the proposed speech enhancement algorithm, a low computational voice activity detection and acoustic noise classification algorithm have also been developed. All these algorithms are promising contributions to the advancement of hearing prostheses, providing a more positive experience to its users.
Exploitation Route As aforementioned, the outcomes of this funding could lead to the advancement of hearing prostheses. Compact speech enhancement algorithms such as the one developed as part of the work funded through this award could also be extended for use in other related areas such as speaker separation and voice recognition. Additionally, it could be implemented in mobile telecommunication devices and in automatic speech recognition algorithms which are now found in many smart home devices. The techniques employed to reduce the computational complexity and increase the compactness of the deep-learning structure utilized in the speech enhancement could be brought forward to other deep-learning-based applications.
Sectors Digital/Communication/Information Technologies (including Software),Electronics,Healthcare,Pharmaceuticals and Medical Biotechnology