Enhanced Hearing Aid Performance with an African Buffalo Optimization-Based Frequency Response Masking Filter

Hearing aid devices primarily serve two key functions: masking and generating passband frames. Fine-tuning these parameters has led to significant improvements in system performance. Amir et al. [17] introduced a 2-level Audio Wave Decompression (AWD) framework within a reconfigurable device to enhance functionality and auditory performance. This design was hardware-validated and succeeded in reducing hardware complexity, although it exhibited a considerable range of matching errors. Amir et al. [18] developed a mono-reconfigurable structure catering to various hearing impairments across patients, with FPGA-based implementation results indicating low power consumption and reduced complexity. The filter parameters were refined using an optimization model.

Shrivastav and Kolte [19] proposed a reconfigurable Frequency Response Masking (FRM) filter bank, optimized with a heuristic technique, yielding over a 60% improvement in auditory compensation. Wei et al. [20] introduced a non-uniform filter bank (NF) model with 16 bands for hearing aid systems, requiring only three multipliers for executing filter band operations. This model demonstrated a lower delay ratio, though it was complex in design. Shrivastav et al. [21] designed a reconfigurable FRM filter utilizing three decomposition levels and optimized it with a swarm-based algorithm. Fan et al. [22] created a deep learning-based application for hearing aids capable of automatically recognizing five types of audio effects, achieving a 98.8% accuracy rate with thirty audio blocks for training, albeit with a longer execution time.

Zhang et al. [23] designed a unified modulated filter bank aiming to reduce complexity by minimizing the use of multipliers, but this approach resulted in lower hardware accuracy. Tiwari et al. [24] presented a sliding-band compression model for sensorineural hearing loss, offering a distinct amplitude range for various compression ratios but with high processing times that affected performance. Zheng et al. [25] introduced an FRM-based non-uniform filter bank for hearing aids, achieving desired stopband attenuation with optimal word length for filter coefficients, which simplified hardware design. Bhuyan et al. [26] proposed an area- and power-efficient reconfigurable filter using interpolated subband distribution and a LUT-less DA method. Devis and Manuel [27] developed a cost-effective three-level filter bank from a single FIR filter using the Parks McClellan algorithm and fractional interpolation, aiming for minimal hardware requirements. Lastly, Parameshappa and Jayadevapp [28] proposed an FRM-based uniform digital filter bank utilizing an interpolated half-band FIR filter, which achieved an acceptable range of matching error and computational complexity.

This extensive review of existing reconfigurable filter designs underscores a significant opportunity for enhancements in matching error, delay, power consumption, and hardware complexity. The primary aim of the proposed research is to address these gaps and provide hearing aid users with a more effective and user-friendly experience. The proposed approach begins with the input of the original speech analog signal, followed by the development of a novel African buffalo-based Frequency Response Masking Reconfigurable Filter (AB-FRMRF) using MATLAB and the Xilinx tool. The filter bank's characteristics, such as matching error, delay, frequency response, area, and power consumption, are meticulously optimized using the African buffalo optimization algorithm. The proposed model is then thoroughly evaluated against existing literature on parameters including delay, matching error, area, and power, ensuring a comprehensive analysis of its efficacy.

Typically, digital hearing aids are effective for people with hearing loss. However, the existing hearing aids have uncomfortable noise, high delays and more matching errors. The proposed study article aims to model a novel African buffalo-based Frequency Response Masking Reconfigurable Filter design (AB-FRMRF) to enhance the functionality of hearing aid devices to improve the hearing ability of hard-of-hearing people. Additionally, the African buffalo fitness function optimizes the parameters of the designed filter bank. The proposed filter bank’s masking parameter also covers the full-range frequency signal. Figure 2 describes in detail how the proposed AB-FRMRF system operates.

2.png

Figure 2. Methodology for AB-FRMRF design

The A/D converter is given the input audio signal. The proposed filter then processes the input to increase the effectiveness of the hearing device. The D/A is then fed the filter bank’s output, and the parameters are determined. As a result, the designed approach has produced the best results while using less power and experiencing less delay.

4.1 Design of FRM reconfigurable filter model

3.png

Figure 3. Proposed AB-FRMRF structure

The proposed AB-FRMRF model is developed to augment hearing aid performance. The proposed optimized design has decreased complexity and delay while achieving minimal matching error. This work builds the AB-FRMRF filter and optimizes the filter bank parameters using the African Buffalo Optimisation (ABO) [30] methodology. The FRMRF filter is hybridized with ABO to improve the filter bank characteristics. The structure of AB-FRMRF filters is paired with masking filters, the AB fitness function, and periodic model filters for achieving a narrow transition band using FIR filters with linear phase response. In this method, the proposed filter divides the entire frequency range into separate zones with variable bandwidth, utilizing the masking stage. The masking filters are represented as Fm (z) and Fmc (z). Also, the periodic model filters Ga(z) and G_a(z^m) are designed by exchanging every delay component of the narrow band filter into (m) delay components by fitness (k), which are depicted in Figure 3.

Additionally, passband and stopband edges in the filter are mentioned as $\phi$ and $\theta$. The matching Filter's transition width $\mathrm{G}\left(\mathrm{z}^{\mathrm{m}}\right)$ is $(\theta-\phi) / m$. Also, these filters are mentioned as $\mathrm{G}_{\mathrm{c}}\left(\mathrm{Z}^{\mathrm{m}}\right)$ and expressed in Eq. (6) as:

$G_c\left(z^m\right)=\left(z^{\frac{-N-1}{2} k}-G\left(z^m\right)\right)$             (6)

where, N denotes the impulse response length of G(z^m), subsequently, the masking filters of this structure have cascaded to G(z^m) and G_c(z^m) and are combined to create an FRM filter. As a result, Eq. (5) is used to represent the developed AB-FRMRF filter’s transfer function.

The proposed AB-FRMRF filter’s frequency response is mentioned in Eq. (7).

$G_a(z)=\mathrm{G}(\mathrm{Zm}) \mathrm{Fm}(\mathrm{Z})+\left(z^{\frac{-N-1}{2} k}-G\left(z^m\right)\right) \mathrm{Fmc}(\mathrm{Z})$              (7)

The proposed AB-FRMRF filter also has group delay, as mentioned in Eq. (8).

$D=\frac{(N-1) m}{2}+d$              (8)

where, d = max((N_a-1)/2, (N_c-1)/2).

4.2 Methodology for optimized AB-FRMRF model design

The proposed AB-FRMRF model’s methodology is described in Figure 4.

4.png

Figure 4. Flow chart of proposed AB-FRMRF algorithm

Here, the constructed filter bank’s parameters, such as delay and matching error, are further tuned using the AB fitness function indicated in Eq. (9):

k = m.a + r₁(bg_max – w.a) + r₂(bp_max - w.a)                      (9)

where, r₁ and r₂ are the learning parameters for identifying the filter bank parameters, whereas m.a and w.a denote the actual loss parameters, bg_max represents the matching error optimizing factor, and bp_max is the optimizing factor for group delay. As a result, outcomes like matching errors and delays were more effective. More sub-bands are generated because of the reconfigurable nature of the proposed filter-bank model.

4.3 Speech perception analysis

Speech perception of a person using a hearing aid is the ability to read, interpret, and understand language sounds. The input signal from the speaker includes several environmental effects like background noise, reverberation, and signal level reduction due to the distance between the speaker and the hearer. Speech perception consists of the hearer’s outcome of finding an acceptable fit between the emotional input provided by the speaker and the linguistic structure intended by the speaker.

The suggested filter bank evaluates each patient’s hearing loss and speech perception. Moreover, to perform speech perception analysis, various categories of input signals are given to the filter bank to identify the ability to understand the sound signals.

Table 1. Different audio signal dataset

Note: ¹Hertz

A novel filter bank is presented in this paper to augment the performance of the HA system. The Xilinx/Kintex and MATLAB frameworks are used to develop the proposed AB-FRMRF replica. The FPGA device model is XC7K325T-2FFG900C, and parameters, including matching error, delay, frequency, area, and power, are determined. Additionally, the obtained findings are contrasted with other widely used techniques for performance parameters, viz., matching error, frequency, delay, power, and area. Table 1 displays the various audio signal datasets that are used.

Initially, the proposed filter is designed mathematically in the MATLAB environment. The filter’s performance was analyzed by evaluating key metrics like matching error and delay. Hereafter, the speech signal is converted to digital form and trained as input for the Xilinx FPGA model. Here, the novelty in the FPGA module is Brent Kung’s adder. It used less area to run the process. The RTL schematic model of the proposed system designed for Vivado 2018 is depicted in Figure 8. After FPGA synthesis, the filter coefficients are again imported into MATLAB to optimize the filter coefficients. The optimized filter coefficients are then given to Xilinx, and the optimized results are validated. The proposed filter reports a delay of 4.443 ns in the FPGA module against the typical reconfigurable filter delay, which was high up to 6.008 ns in the literature.

The Simulation results of the proposed filter design are shown in Figure 9.

8.png

Figure 8. RTL schematic model of the proposed filter

9.jpg

Figure 9. Simulation result of the proposed design

6.1 Validation

Hearing loss can cause significant harm to human life, relationships, work, and physical and emotional well-being, making it one of the most critical health problems affecting human life. In this research, the novel AB-FRMRF filter is designed to enhance the efficiency of hearing aids to improve the hearing ability of hearing loss patients. This scheme includes digital filters, which tune the amplitude based on the patterns of hearing loss (HL). The hearing loss types are also mentioned as usual: mild, moderate, and severe. Moreover, the developed filter separates the sound signals into various subbands. Selective gain has been provided to every subband to match the audiogram of the particular person.

The average hearing range of humans is 0–20 dB. The threshold range of the hearing loss levels is mentioned in Table 2. In this work, the hearing ability of hearing loss patients is improved using a hearing aid system designed using a novel AB-FRMRF model. The audio signals are initially taken for the system’s input and are collected under various environmental conditions, like a human voice and music. If the distance between the hearing aid and the source is considerable, the noise and errors in the audio signal are high. Eight sub-bands of input were separated using the specified filter. Specifications of the designed filter bank are shown in Table 3.

Table 2. Threshold range of hearing loss

Notes: ¹Decibel

Table 3. Specification of the designed filter

Note: ¹KiloHertz

Only thirteen multipliers are used when the developed filter satisfies these requirements. Finally, the hearing aid system employs the designed filter to compute the performance metrics. With the aid of this AB-FRMRF model, the hearing aid system’s effectiveness is increased while matching error, frequency, power, and delay are all optimized.

6.2 Performance metrics

The proposed AB-FRMRF approach effectively designs the filter bank and tunes the parameters such as matching error, delay, frequency, power, and area. Initially, the audiogram for various HL categories, such as mild, moderate, mild to severe, and profound HL for left and right ears, is determined.

An audiogram is a graph representing the patient’s hearing ability based on various frequencies. A consolidated representation of a set of sample audiograms of varying hearing loss profiles is presented in Figure 10. Additionally, different hearing loss categories are considered for audiogram matching. The subbands are formed during the breakdown of an input signal. Based on the hearing capability of each person, amplification is provided in every subband using the gain block.

10.jpg

Figure 10. Sample audiogram for various types of HL

The number of bands, frequency range, bandwidth, and gain for each audiogram are analyzed, and these design parameters are given in Tables 4-7, respectively.

Table 4. Design parameters of AB-FRMRF for mild HL

Note: ¹Subbands

Table 5. Design parameters of AB-FRMRF for moderate HL

Note: ¹Different frequencies

Table 6. Design parameters of AB-FRMRF for severe HL

Note: ¹Different bandwidth

Table 7. Design parameters of AB-FRMRF for profound HL

Note: ¹Decibel

Additionally, the magnitude responses of the utilized filters are represented graphically. The periodic model, masking, and narrowband low-pass filters employed in the proposed FRMRF produce the optimal output signal. Figure 11 to Figure 15 display the obtained audiogram for various hearing impaired with mild to profound hearing loss.

11.png

Figure 11. Audiogram fitting for mild HL

12.png

Figure 12. Audiogram fitting for mild to moderate HL

13.png

Figure 13. Audiogram fitting for moderate to severe HL

14.png

Figure 14. Audiogram fitting for severe HL

15.png

Figure 15. Audiogram fitting for profound HL

6.2.1 Matching error

Matching error (ME) is the measure of variation between the audiogram curves that are matching and the hearing loss curves. The matching error of the proposed approach is compared with existing methods and is shown in Figure 16.

16.png

Figure 16. Comparison of ME curves for various methods

The proposed model is compared with existing models like the Signed Approximate Multiplier FRM (SAM-FRM) filter [31], AWD [15], and Non-uniform Filter Bank (NF) approach [18] for evaluating the performance of the proposed model. The comparison of matching errors for different HLs is detailed in Table 8.

Table 8. Comparison of matching error in different methods

Note: ¹Hearing loss types

6.2.2 Delay calculation

The delay calculation is an essential parameter in the HA filter bank. The delay value should be low for a better filter bank system. Generally, non-uniform filter banks have more delay than uniform filters. High delays can affect the lip-reading ability of hearing loss patients. The proposed filter design has attained a lower delay value when compared with other existing methods detailed in Table 9.

The calculation of the delay value of the proposed HA design is compared with existing methods like SAM-FRM, AWD, and NF. Here, the proposed filter model utilized the fitness function of African buffalo to optimize the delay value.

Table 9. Comparison of delay

17.png

Figure 17. Comparison of delay

The proposed filter bank model has optimized the delay for various hearing loss types at multiple frequencies. The delay value of the proposed approach is compared with the existing techniques and is presented in Figure 17.

In our proposed non-uniform filter design with three sub band distribution schemes, the use of FRM technique further optimized with ABO algorithm in hybrid combination with distributed arithmetic method resulted in variable customized non-uniform sub band distribution with narrow transition bandwidth and optimized filter coefficients leading to improved matching error and delay as against the SAM-FRM [31] and AWD [15] based designs.

Table 10. Comparison of power

Note: ¹Megahertz

6.2.3 Power analysis

Power is vital for all digital applications; better options exist if the finest technique consumes more power. Power analysis is performed in Xilinx Vivado for the target Kintex 7 FPGA family. The comparison of power analysis is described in Table 10, and its statistical comparison is depicted in Table 11. Results show that the proposed design consumes less power at all frequencies than the conventional reconfigurable filter proposed in the study [32].

In our proposed methodology, using the Distributed Arithmetic technique combined with employing Brent Kung’s parallel prefix adder for adding DRAM partial products instead of the conventional pipeline adder tree has significantly improved power requirements. It is thus apt for small-size hearing aids with limited battery life.

Table 11. Statistical comparison of power consumption