Comparison of IIR filterbanks and FFT filterbanks in cochlear implant speech processing strategies

Abstract EN

The first stage of all cochlear implant speech processing strategies is a filter bank which decomposes the speech signal into several frequency bands.

Two types of filter banks can be used to achieve this goal and mimic the spectral analysis performed within the cochlea.

The first is based on the Fast Fourier Transform (FFT).

The second type of filter bank is composed of Infinite Impulse Response (IIR) filters.

This study aims to compare the performances of the two filter banks by evaluating the outputs of acoustic models.

Therefore, a standard strategy was implemented and tested with four different filter banks.

Butterworth IIR filters were used in the first and the FFT with 128, 256 and 512 points were used in the three other filter banks.

The evaluation of the performances was obtained by considering the word recognition obtained by an Automatic Speech Recognition (ASR) system developed within this work and the accuracy of the pitch coding.

The results showed that the word recognition was higher when using the IIR filter bank or the 256-point FFT.

Besides, increasing the number of FFT points gave more accurate pitch contours.

However, the pitch coding performance was better with the IIR and the 512-FFT filter banks.