EE Seminar: Real-Time Cochlear Noise Reduction Algorithm for Embedded Platforms
Speaker: Uriya Saroussi
M.Sc. student under the supervision of Prof. Miriam Furst-Yust and Prof. Shlomo Weiss
Wednesday, September 14th, 2016 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Real-Time Cochlear Noise Reduction Algorithm for Embedded Platforms
Hearing impaired (HI) persons often struggle with understanding speech in the presence of loud background noise even when wearing hearing assistive devices. The vast majority of noise reduction algorithms today are multiple-channel algorithms that use the correlation between the inputs to differentiate the speech from the noise. While significantly reducing background noise, these algorithms still result in a noisy signal that can affect hearing impaired persons ability to understand speech. The study of Fink et al. (2012) showed that HI people have difficulties in understanding speech even with relatively high SNRs. Applying a single-channel noise reduction algorithm on the signal resulted by the multi-channel algorithm would improve the speech intelligibility of HI persons.
The Cochlear Noise Reduction Algorithm (CNRA) is a single-channel binary mask algorithm which was tested on a group of HI subjects wearing their hearing assistive devices and was proven to significantly improve speech intelligibility. It is based upon the one-dimensional cochlear model with embedded outer hair cells (OHC) developed by Cohen and Furst in 2004. The output of the model is used to differentiate the speech from the noise, thus improving the SNR of the noisy input signal. Massive computations are required to solve the model on which the CNRA depends, making its use in real-time applications a challenge.
It is the purpose of this study to design a variant of the CNR algorithm which could perform in real-time on portable platforms, without damaging the quality of the output signal, so that it can be used as the basic building block of a hearing-aid device.
A new variant of the CNR algorithm which is able to run in real-time on an Android based smartphone and still produces near identical results to previous CNRA variants was designed.
Another important objective is that the new CNRA variant would run upon hardware that most people carry on a daily basis, thus obviating the need of the HI population to carry additional devices. Such a device that complies with the hardware requirements and is also considered a commodity nowadays is the smartphone. Therefore, an Android hearing-aid application which acquires input samples from the device's microphone, applies the CNRA on the recorded samples and outputs them using the device's output interface was designed and implemented. The newly designed CNRA variant was also implemented in such a way that would allow to run it efficiently on a smartphone and to comply with the demands of the Android audio recording and playing systems and to be integrated into the real-time hearing-aid application.
The outputs of the newly designed CNRA variant running on the Android-based smartphone were compared to outputs from a previous non real-time CNRA variant implemented on a PC (the latter presents a significant speech intelligibility improvement for HI subjects wearing their hearing-aid devices). The comparison yielded nearly identical results.
