FPGA-Based Comb Filters for Binaural Hearing Aids

1. FPGA-Based Implementation of Comb Filters Using Sequential Multiply-Accumulate Operations for Use in Binaural Hearing Aids by S. G. Kambalimath, P. C. Pandey P. N. Kulkarni, Mahant-Shetti and S. G. Hiremath

2. PRESENTATION OUTLINE 1 Introduction 2 Implementation 3 Results 4 Conclusion

3. Intro.1/5 1. INTRODUCTION Sensorineural hearing loss ●Elevated hearing thresholds ●Decreased dynamic range and abnormal loudness growth ●Increased temporal & spectral masking

4. Intro.2/5 Binaural Dichotic Presentation Using Complementary Comb Filters for Persons Using Binaural Hearing Aids [Lyregaard, 1982; Lunner et al., 1983; Kulkarni et al, 2012] Spectral components likely to mask or get masked by each other are presented to different ears, for reducing the adverse effect of increased intraspeech spectral masking and improving speech perception. A binaural hearing aid using comb filters

5. Intro.3/5 Auditory Critical Band Based Perceptually Balanced Comb Filter Pair [Kulkarni et el., 2012] Filter responses: magnitude responses designed for perceptually balanced loudness, linear phase responses. Magnitude responses of comb filter pair with floating-point coefficients & offline implementation. Results of test on hearing impaired listeners: (i) 14 – 31% increase in consonant recognition, (ii) 0.26 s decrease in response time, (iii) no adverse effect on the localization of broadband sounds.

6. Intro.4/5 Wide-band spectrograms (Δf = 0.3 kHz) of a sentence ‘would you write gun’: (a) diotic, (b) dichotic, processed with a pair of comb filters [Kulkarni et el., 2012]

7. Intro.5/5 Research Objective Efficient FPGA basedImplementation of comb filter pair, as the first step for its use in binaural hearing aids, using different architectures.

8. Imp.1/4 2. FPGA BASED IMPLEMENTATION FPGA and Audio codec interfacing on FPGA board used for comb filter implementation Comb filter realization as direct-form FIR filter structure.

9. Imp.2/4 Comb filter realization as transposed-form linear phase FIR filter structure. FPGA-based implementation of direct-form FIR filter structure using parallel multiply-accumulate operations. ▪ Uses (N-1) delays, N multipliers and, (N-1) adders

10. Imp.3/4 FPGA-based implementation of transposed-form linear phase filter structure using parallel multiply-accumulate operation. ▪Explots symmetry in filter coefficients. ▪No. of multipliers reduced to half as needed in direct form implementation.

11. Imp.4/4 FPGA-based implementation of direct-form FIR filter structure using sequential multiply-accumulate operations. ▪ Uses only one adder and one multiplier. ▪ Needs extra resources for implementing Multiplexer, control logic, and clock generator.

12. Result.1/5 3. RESULTS All the implementations worked satisfactorily for sampling frequency of 10 kHz. Informal listening tests: Binaural presentation of the processed test stimuli showed no perceptual distortion for the processed sounds indicating nearly perfect perceptual fusion of the binaural sounds Magnitude responses of the 257-coefficient comb filters (dark & light for L & R filters) using Direct-form parallel implementation.

13. Result.2/5 Magnitude responses of the 257-coefficient comb filters (dark & light for L & R filters) using transposed form parallel implementation. Magnitude responses of the 257-coefficient comb filters (dark & light for L & R filters) using direct form, sequential implementation.

14. Result.3/5 Magnitude responses of the 513-coefficient comb filters (dark & light for L & R filters) using Direct-form parallel implementation. Magnitude responses of the 513-coefficient comb filters (dark & light for L & R filters) using transposed form parallel implementation.

15. Result.4/5 Magnitude responses of the 513-coefficient comb filters (dark & light for L & R filters) using direct form, sequential implementation. All the filter responses have pass-band ripple below 2 dB and cross-over gains of −4 to −8 dB, stop-band attenuation is greater than 25 dB for 513 coefficient filters and greater than 18 dB for 257-coefficient filter. Deviation in the sum of magnitude responses on a linear scale is below 0.098 for all responses. The properties of the 513-coefficient FPGA- based comb filters closely match with those reported earlier (Kulkarni et al., 2012)

16. Result.5/5

17. Concl.1/1 3. CONCLUSIONS All the implementations worked satisfactorily for sampling frequency of 10 kHz. Implementation using a 16-bit codec, 15-bit signed coefficients, and 32-bit registers resulted in satisfactory filter responses and used only a fraction of resources available on the chip. The filter architecture using sequential multiply-accumulate found more efficient in resource utilization with the scope for implementing other processing blocks of a hearing aid on the same chip. Implementation of prototype hearing aid with dynamic range compression, frequency-selective response, and comb filters can be taken up for developing a hearing aid. FPGA design can be converted to an ASIC for developing wearable hearing aid.

18. THANK YOU