A Design Method for Gammachirp Filterbank for Loudness Compensation in Hearing Aids

Abstract

Because the hearing impaired often experience different degrees of hearing loss along with the loss of frequencies, the loudness compensation algorithm in hearing aids decomposes the speech signal and compensates with different frequency bands based on their audiograms. However, the speech quality of the compensated signal is unsatisfactory because the traditional filterbanks fail to fully consider the characteristics of human hearing and personalized hearing loss. In this study, an effective design for the gammachirp filterbank for the loudness compensation algorithm was proposed to improve the speech quality of hearing aids. Firstly, a multichannel gammachirp filterbank was employed to decompose the signals. Then, the adjacent bands were merged into one channel, guided by the proposed combination method. After obtaining the personalized filterbank, each band conducted a loudness compensation to match the requirements of the audiograms. The excellent advantage of the gammachirp filterbank is that it can simulate the characteristics of the basilar membrane. Furthermore, the novel channel combination method considers the information from the audiograms and the relationship between frequency ranges and speech intelligibility. The experimental results showed that the proposed multichannel gammachirp filterbank achieves better speech signal decomposition and synthesis, and good performance can be gained with fewer channels. The loudness compensation algorithm based on the gammachirp filterbank effectively improves sentence intelligibility. The sentence recognition rate of the proposed method is higher than that of a system with a gammatone filterbank by approximately 13%.

1. Introduction

As the population ages, the proportion of the population that experiences hearing loss is constantly expanding, and more patients are choosing to wear digital hearing aids to improve their hearing quality [1]. The processing of speech signals in hearing aids mainly includes loudness compensation, noise reduction, and adaptive echo suppression [2,3,4]. The loudness compensation algorithm compensates with different frequency bands based on the audiogram of the hearing impaired to improve speech intelligibility. Therefore, loudness compensation, as one of the key algorithms, is the main task of digital hearing aids. This can be implemented by first decomposing the speech into subbands using a filterbank and then selectively amplifying the subband signals according to different intensities to bring the hearing level of the hearing impaired closer to normal hearing. The channel partition and filterbank design are the key parts of the compensation algorithm, as these directly affect the performance of the hearing aid [5].

The classic filterbank used in hearing aids decomposes speech signals into uniform subbands. For example, the filterbank in [6] was based on discrete Fourier transformation and the filterbank in references [7,8] was developed by employing interpolated, complementary linear-phase FIR filters and a frequency–response masking technique. Dividing the spectral into uniform bands is the most straightforward method, but does not consider the characteristics of human hearing and the differences between audiograms. Thus, nonuniform filterbanks that can adapt to the hearing characteristics and the matching requirements of different audiograms have recently attracted widespread attention from scholars [9,10,11].

The gammatone filterbank is one of the representative nonuniform filterbanks used in hearing aid applications that is able to simulate the auditory model of the cochlea of the human inner ear. It can also simulate the spectrum analysis functions, and the frequency selectivity of the human ear [12,13]. However, the amplitude–frequency response curve of the gammatone filter is symmetrical at about the center frequency and the amplitude is independent of the intensity, which cannot reflect the asymmetry and intensity-related characteristics of the basilar membrane curve. Considering these drawbacks, some scholars have put forward a model of the gammachirp filter, developed on the basis of the gammatone filter in accordance with physiological and psychological experimental data. This filter retains the advantages of the gammatone filter, but also makes up for its defects in terms of asymmetry and intensity dependence [14]. In recent years, the gammachirp has made some achievements as an ideal auditory filter in speech signal processing, such as audio coding, formant estimation, speaker recognition, speech signal analysis, and simulation of the human ear basement membrane [15,16,17,18,19].

To address these challenges, this study proposes a novel method to apply the gammachirp filterbanks to decompose input signals. Furthermore, we propose a new channel-division method to satisfy different types of audiogram matching requirements. The key idea of the proposed method is to consider the relation between frequency ranges and speech intelligibility and the specific audiogram features. The proposed multichannel filterbank was evaluated by two objective indices: perceptual evaluation of speech quality (PESQ) [20,21] and short-time objective intelligibility (STOI) [22]. The results show that the gammachirp filterbank has good reconstruction performance, and the speech quality index of the processed signal is higher than that of the gammatone filterbank. From the subjective assessment experiments, it was found that the loudness compensation algorithm based on the gammachirp filterbank effectively improves the sentence intelligibility. The compensated signals also obtain higher speech evaluation indices for the hearing aid: hearing-aid speech perception index (HASPI) and hearing-aid speech quality index (HASQI) [23,24].

The main contributions of this study are as follows:

• We applied the gammachirp filterbank as a component in the loudness compensation system. The filter parameters were calculated based on the human auditory characteristics and the intensity of the input signal. The obtained multichannel gammachirp filterbanks model the frequency selectivity characteristics of the human ear;
• We introduced a new method of channel combination to reduce the computational complexity. The adjacent channels were merged by the characteristics of the special audiogram and the relationship between frequency ranges and speech intelligibility. After obtaining the personalized filterbank, each band carried out loudness compensation, based on the curve provided by the hearing-impaired patients;
• We conducted an objective evaluation and subjective assessment experiments. The results illustrate that the proposed method can considerably improve sentence intelligibility. The input speech can be reconstructed with fewer channels and satisfy different audiogram matching requirements.

2. Design of Gammachirp Filterbank

The basic structure of the M-channel filterbank applied in loudness compensation is shown in Figure 1. The input signal $x(n)$ is first decomposed into different frequency bands to obtain subband signals. The $G_k(z),\left\{k=0,1,\dots ,M-1\right\}$ represents the analysis filterbank, which carries out subband decomposition. Then, each band carries out loudness compensation in a specific frequency range. Some prescriptive procedures are used to set the amplifying characteristics of the hearing aids, such as NAL-NL2 [25], DSLv5 [26], and FIG6 [27], which are also called prescription formulas. The subsignal can be modified according to the required gain to satisfy the hearing comfort requirements of diverse patients. Lastly, the processed subsignals are combined into a reconstructed signal as the final output under the function of the synthesis filterbank $F_k(z),\left\{k=0,1,\dots ,M-1\right\}$ [28].

The gammachirp filter is a kind of nonlinear filter, in accordance with the nonlinear response of the human ear basilar membrane to frequency [29], and its temporal impulse response function is as follows:

$$g(t)=B^n{t}^{n-1}{e}^{-2\pi Bt}\cos (2\pi f_{r}t+c\ln t+\psi )$$

(1)

where $c$ is the rate of the frequency modulation, which is called the chirp factor with a value in the range of [−3, 3], $\ln t$ is the natural logarithm of time, and $\psi$ is the initial phase. The asymmetric parameter $f_r$ is the center frequency of the gammachirp filter. Compared with the gammatone, the impulse response of gammachirp adds the $c\ln t$. The $f_r$ changes with the value of $c$, and depends, to some extent, on the parameter $B$ and the order $n$. The value of parameter $B$ satisfies the equation $B=b·ERB(f_r)$ and $ERB(f_r)$ is the equivalent rectangular bandwidth of an auditory filter at $f_r$. Typically, $n=4,b=1.109$ [30]. Additionally, the chirp parameter c is linearly related to the intensity of the stimulus signal $P_s$ (in dB SPL), which can be calculated by Equation (2).

$$c=3.38-0.107P_s$$

(2)

When $P_s$ is greater than 30 dB SPL, the chirp factor $c$ becomes negative, and the slope of the lower skirt of the frequency response changes more slowly than that of the upper skirt, which means that the change in the slope is steeper in the high frequency [30,31]. Conversely, the situation is reversed if $c>0$.

The amplitude–frequency response of the gammachirp can be obtained by a Fourier transform of Equation (1) [32]:

$$\begin{array}{ll}{G}_C(f)& =\frac{a\mathsf{\Gamma }(n+jc)e^{j\varphi }}{{\left\{2\pi bERB(f_r)+j2\pi (f-f_r)\right\}}^{n+jc}}=\frac{a\mathsf{\Gamma }(n+jc)e^{j\varphi }}{{\left\{2\pi \sqrt{{\overline{b}}^2+{(f-f_r)}^2}\cdot e^{j\theta }\right\}}^{n+jc}}\\ & =\overline{a}\frac{1}{{\left\{2\pi \sqrt{{\overline{b}}^2+{(f-f_r)}^2}\right\}}^n\cdot e^{jn\theta }}\frac{1}{{\left\{2\pi \sqrt{{\overline{b}}^2+{(f-f_r)}^2}\right\}}^{jc}·e^{-c\theta }}\\ & \theta =\arctan \frac{f-f_r}{\overline{b}}{,}_{}^{}{}_{}^{}\overline{b}=bERB(f_r){,}_{}^{}{}_{}^{}\overline{a}=a\mathsf{\Gamma }(n+jc)e^{j\varphi }\end{array}$$

(3)

where the first term $\overline{a}$ is a constant, and the second multiplier term is the amplitude–frequency response of the gammatone filter, which is denoted by $G_T(f)$, while the last multiplier term represents the asymmetric function $H_A(f)$. After amplitude normalization, Equation (3) can be written as:

$$G_C(f)=G_T(f)·H_A(f)$$

(4)

The corresponding amplitude spectrum can be expressed as follows:

$$\left|G_C(f)\right|=\frac{\overline{a}}{{\left\{2\pi \sqrt{{\overline{b}}^2+{(f-f_r)}^2}\right\}}^n}·e^{c\theta (f)}=\left|G_T(f)\right|·\left|H_A(f)\right|$$

(5)

It can be seen from previous derivations that the gammachirp filter can be realized by cascading an asymmetric filter $H_A(f)$ with a gammatone filter $G_T(f)$. Figure 2 provides a block diagram of the signal flow in the gammachirp filter.

The implementation of the asymmetric filter in the gammachirp takes advantage of the method described in [31], and the IIR asymmetric compensation filter is defined as follows:

$$H_A(f)\approx H_C(z),z=e^{j2\pi f/f_s}$$

(6)

$$H_C(z)={\displaystyle \prod _{k=1}^N{H}_{Ck}(z)}$$

(7)

$$H_{Ck}(z)=\frac{(1-r_k{e}^{j{\phi }_k}{z}^{-1})(1-r_k{e}^{-j{\phi }_k}{z}^{-1})}{(1-r_k{e}^{j{\varphi }_k}{z}^{-1})(1-r_k{e}^{-j{\varphi }_k}{z}^{-1})}$$

(8)

The $f_s$ is the sampling frequency, and $N$ is the total number of responses in the cascade, which is usually taken as 4 [31]. The $k\text{-}\mathrm{th}$ pole and zero are defined with the absolute value $r_k$:

$$r_k=\exp (-p_1{(p_0/p_3)}^{k-1}·2\pi bERB(f_r)/f_s)$$

(9)

and the phase of the pole ${\varphi }_k$, and zero ${\phi }_k$ is:

$${\varphi }_k=\{\begin{array}{c}2\pi \{f_r+\Delta f_r\}/f_s,f_r+\Delta f_r\ge 0\\ 0,otherwise\end{array}$$

(10)

$${\phi }_k=\{\begin{array}{c}2\pi \{f_r-\Delta f_r\}/f_s,f_r-\Delta f_r\ge 0\\ 0,otherwise\end{array}$$

(11)

$$\Delta f_r={(p_0/p_3)}^{k-1}·p_2·c·bERB(f_r)$$

(12)

where $p_0,p_1,p_2$ and $p_3$ are a set of positive coefficients related to the chirp factor.

The final filterbank is composed of a series of gammachirp filters with various center frequencies, and the characteristic frequencies of the basilar membrane are usually selected as the center frequency of each subfilter [33,34]. The center frequency of a filter in the filterbank can be calculated according to (13) [35]:

$${\widehat{f}}_{\mathrm{i}}=\frac{1960·i+1038.8}{26.28-i},i=1,2,3,\dots$$

(13)

where ${\widehat{f}}_{\mathrm{i}}$ represents the center frequency of the $i\mathrm{th}$ channel. Thus, a multichannel filterbank with an unequal width can be derived to match the sensitivity of human hearing.

3. Channel Division Method

Some studies have shown that information in distinct frequency bands contributes differently to speech intelligibility, and the energy of voiceless consonants in Chinese is mainly concentrated in the medium-high frequency band [36,37]. The loss of sensitivity and resolution in the medium-high frequencies of the hearing impaired will cause difficulties in speech recognition. The relation between frequency ranges and speech intelligibility is shown in

Table 1. Relation between frequency range and intelligibility.

Frequency Range (Hz)	Intelligibility (%)
0–250	2
250–500	3
500–1000	35
1000–2000	35
2000–4000	13
4000–8000	12

[36].

A new method of channel combination that considers the contribution of frequency components for speech intelligibility is proposed in this section. Firstly, the hearing threshold values at the center frequencies of each channel must be obtained according to the audiogram of the hearing-impaired patient by linear interpolation. Then, the difference between hearing thresholds at the center frequency of adjacent subbands is compared with a set threshold. If the difference is less than the set merging threshold, the two channels can be directly merged into a new channel. Next, the combination thresholds are determined with the influence of each frequency range on speech intelligibility. The combination threshold of the frequency bands with large contributions to intelligibility is 5 dB, which means that channel combining can be performed when the hearing threshold difference at the center frequency of adjacent channels in the 500–2000 Hz band is less than or equal to 5 dB. Similarly, the threshold can be set according to the intelligibility contribution of other frequency bands. In this study, the threshold for the 0–500 Hz band is 20 dB, and 10 dB for the 2–8 kHz range. This channel division method not only meets the matching requirements of audiograms to ensure the performance, but also reduces the number of subband filters and the computation amount.

Consider the transfer function of the M-channel analysis filters $G_{Ck}(z),\left\{k=0,1,\dots ,M-1\right\}$, which are gained by merging $l_i$ adjacent analysis filters.

$${\tilde{G}}_{Ci}(z)={\displaystyle \sum _{k=n_i}^{n_i+l_i-1}{G}_{Ck}(z)}{,}_{}^{}i=0,1,\dots ,\widehat{M}-1$$

(14)

Here, $l_i$ is the total number of channels to be combined, $n_i$ is the upper band-edge frequency $(n_0=0<n_1<n_2<\dots <n_{\widehat{M}}=m=M)$, and $\widehat{M}$ is the channel number of the new gammachirp filterbank. The synthesis filter ${\tilde{F}}_{Ci}(z)$ can be gained using a similar method.

$${\tilde{F}}_{Ci}(z)=\frac{1}{l_i}{\displaystyle \sum _{k=n_i}^{n_i+l_i-1}{F}_{Ck}(z)}{,}_{}^{}i=0,1,\dots ,\widehat{M}-1$$

(15)

4. Experiments and Discussion

4.1. Effect of Chirp Factor on Gammachirp Filter

The amplitude–frequency response curve of the gammachirp filter at different center frequencies is shown in Figure 3a–c when c = 0, c = 2, and c = −2, respectively. It can be seen that the frequency–domain profile of the gammachirp filter shows obvious asymmetry. There is a peak drift relative to the curve of the gammatone filter (c = 0); that is, the maximum amplitude does not appear at the center frequency. In addition, filters with different center frequencies have distinct bandwidths. These characteristics are consistent with the features of the basilar membrane. On the other hand, the value of the chirp factor reflects the intensity of the input speech. Figure 3d shows the amplitude–frequency response curve of the gammachirp filter when the center frequency is taken as 2 kHz under diverse input speech intensities. It can be observed that as the intensity of the stimulus increases, the gain at the peak of the gammachirp filter also increases.

4.2. Comparison of Speech Quality of the Reconstructed Signals

To objectively evaluate the performance of the gammachirp filterbank, this study first designed a 16-channel gammachirp filterbank, and combined the channels to obtain 8-channel and 4-channel filterbanks according to an age-related hearing-loss audiogram. The speech test signals passed through the gammachirp filterbank and gammatone filterbanks, and we compared the quality of those output speech test signals.

The test speeches were selected from the CASIA Chinese emotional corpus, which was recorded by the Institute of Automation, Chinese Academy of Sciences. It included four professional speakers and five kinds of emotion: angry, happy, sad, surprised, and neutral, with a total of 9600 different pronunciations. We randomly selected a total of 300 recordings from the corpus as experimental data, all of which were recorded by a speaker in a neutral emotional state. The sampling rate of the test signal was 16 kHz, the sampling precision was 16-bit, and the duration of each file was approximately 2 s.

Two objective speech quality indices, PESQ and STOI, were applied to assess the quality of the output signals from the gammachirp and gammatone filterbanks with different numbers of channels, and those indices are compared in this section. STOI and PESQ are both closely related to human auditory perception and are widely used as evaluation criteria of voice quality. A higher value represents the higher intelligibility of the processed speech.

Figure 4 and Figure 5 are the comparisons of the PESQ and STOI speech quality indices of the 8-channel and 4-channel gammatone filterbanks and gammachirp filterbanks. The index values of the different signals are shown on the vertical y-axis, and the serial numbers of the test speeches are shown on the horizontal x-axis. It can be observed from the comparison that the performance of the 8-channel and 4-channel gammachirp filterbanks obtained by considering the patients’ audiogram characteristics is improved, and the PESQ and STOI values are better than for those of the gammatone filterbank. The indices of the 8-channel and 4-channel gammatone filterbanks and gammachirp filterbanks are compared in

Table 2. Comparisons of speech quality indices.

Index	Number of Channels	Gammatone Filterbanks		Gammachirp Filterbanks
		Mean	Variance	Mean	Variance
PESQ	8-channels	3.7532	0.0117	4.0126	0.0118
	4-channels	3.7233	0.0148	4.0742	0.0153
STOI	8-channels	0.9606	3.4580 × 10−5	0.9701	2.603 × 10−5
	4-channels	0.9574	5.3106 × 10−5	0.9604	4.4718 × 10−5

. Compared with the 8-channel filterbanks, the PESQ and STOI of the 4-channel gammachirp filterbanks are close to those of the 8-channel filterbanks, and the output signal has comparable intelligibility. By all of these comparisons, the gammachirp filterbank designed based on patients’ hearing characteristics can better simulate the filtering properties of the human ear, the output speech has higher intelligibility, and the signal can be approximately and accurately reconstructed.

4.3. Loudness Compensation Experiment

The loudness compensation experiment based on the gammachirp filterbank was carried out to verify the performance of the system, and the experimental results are given from two perspectives: objective evaluation and subjective assessment.

4.3.1. Objective Evaluation

A new method to objectively evaluate speech quality was introduced to provide qualitative and quantitative analyses of processed speech. The speech intelligibility index, HASPI, and the speech quality index, HASQI, which represent system performance were calculated by using the function provided by the toolbox. It is worth noting that the input parameters include the vectors of the hearing loss in six audiometric frequencies, which can be acquired in the audiogram. That is, the result is related to the patient’s hearing loss, and a series of objective experiments can be used instead of the subjective tests.

The test audiograms were selected from the Occupational Hearing Loss (OHL) Surveillance Project of the NIOSH program for the Centers for Disease Control and Prevention in the United States. NIOSH is short for the National Institute for Occupational Safety and Health program, which collects worker audiograms across America [38]. In this work, we retrieved a dataset of 300 audiograms from participants, which were obtained using a standard pure tone audiometry protocol. Air conduction thresholds were measured at seven test frequencies: 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz, 4000 Hz, 6000 Hz, and 8000 Hz. The hearing impairment can be divided into four grades: mild, moderate, severe, and profound according to the diagnostic criteria of the WHO [39]. Among the hearing-impaired patients corresponding to the 300 audiograms, 120 had moderate hearing loss, 100 had severe hearing loss, and 80 had profound hearing loss. The speech test signals used in the experiments were also from the CASIA Chinese emotional corpus. We randomly selected 50 speech test signals from the corpus.

We conducted 300 rounds of loudness compensation experiments in this section according to the audiograms of 300 hearing-impaired patients. In the previous analysis, it can be seen that the 4-channel gammachirp filterbank had good fidelity. In this experiment, a 4-channel filterbank was used as an example. The filterbanks used in each round of experiments were combined according to a specific audiogram. For each audiogram, the main experimental steps were as follows. Based on the channel division method described above, a 4-channel gammachirp filterbank was first obtained. Then, the input speech test signal was passed through the filterbank, and the compensation gains were calculated with the FIG6 prescription formula. The FIG6 prescription formula is a prescription procedure that provides a set of rules to calculate the required gain under three input intensities (40, 65, and 90 dB SPL) according to the hearing threshold. This can be used to adjust the gain to the input signals, making speech intelligible and the overall loudness comfortable [27]. Lastly, the processed subsignals were combined into a reconstructed signal as the final output.

According to the experimental results, there were six main useful groups of 4-channel gammachirp filterbanks which were combined from the 16-channel filterbanks according to the patients‘ audiograms. The frequency responses of these are shown in Figure 6. The group of six filterbanks corresponds to 12, 78, 22, 34, 50, and 104 audiograms, respectively. The audiograms of patients with moderate hearing loss were mainly classified into groups 2 and 5, while severe hearing loss mainly corresponds to groups 4 and 5. The hearing impairment of the profound grade generally belongs to groups 2 and 6.

In addition, we compared the average HASPI and HASQI indices of the output signals after the loudness compensation based on the 4-channel gammatone filterbank and gammachirp filterbank. Figure 7 shows a comparisons of the two indices computed using the output signals based on the 4-channel gammachirp filterbanks and gammatone filterbanks. A significance analysis indicated that the difference between the various indices has a statistical significance (p < 0.0001), and the speech intelligibility and quality indices of the system based on the gammachirp filterbank output are better than those of the system based on the gammatone filterbank.

Furthermore, a comparison was made between the HASPI and HASQI indices of the output signals of the 16-channel and 4-channel loudness compensation systems. Figure 8 shows the comparisons of the two indices that were derived using the output signals that were obtained from the 16-channel and 4-channel gammachirp filterbanks. The average index values of the 50 output signals are shown on the vertical y-axis, and the serial numbers of the audiograms are shown on the horizontal x-axis. It can be observed from the two curves that the performance of the 4-channel gammachirp filterbank was better than that of the 16-channel filterbank. Therefore, the effectiveness of the new channel combination method can be verified, considering the hearing characteristics of patients and reducing the number of channels. A significance analysis shows that the HASPI and HASQI of the 4-channel compensation system are better than those of the 16-channel compensation system, and the difference between the various indices has statistical significance (p < 0.0001).

4.3.2. Subjective Assessment

Ten patients with hearing loss were invited to the following experimental tests and all of them had had the experience of wearing hearing aids for more than three months. Due to personal reasons, four of them withdrew from the trial. The subjective assessment experiments were performed on six hearing-impaired patients to prove the increase in intelligibility. The hearing threshold measurements of the patients who did not wear hearing aids was implemented based on simplified pure tone audiometry, and the measured tones were 125, 250, 500, 750, 1000, 1500, 2000, 3000, 4000, 6000, and 8000 Hz. The audiogram results are shown in Figure 9 for both left and right ears.

The patients were required to listen to test speech signals and to reconstruct speech signals after multichannel loudness compensation. The sound pressure level of the test signals was adjusted to 10 dB, 20 dB, 30 dB, 40 dB, 50 dB, 60 dB, and 70 dB, and for every level, 50 original sentences and 50 compensated sentences were tested. The patients were asked to repeat the sentences. When their reproduction was completely correct, the patients were deemed to have correctly identified the sentence.

The average speech-signal recognition rates before and after loudness compensation based on the 4-channel gammachirp and gammatone filterbanks are shown in

Table 3. Comparison of the speech recognition rate.

SPL (dB)	The Recognition Rate of the Original Speech (%)	The Recognition Rate of Compensation Signal Based on Gammatone Filterbank (%)	The Recognition Rate of Compensation Signal Based on Gammachirp Filterbank (%)
10	0	0	0
20	0	0	0
30	0	4	16
40	10	12	22
50	32	38	46
60	58	68	76
70	68	80	88

. The experiments show that after loudness compensation, the patients’ speech recognition rate of speech signals at various sound pressure levels had improved. Additionally, the hearing threshold of the patients decreased, and some speech files with low sound pressure levels could be identified by the patients. When the sound pressure level of the testing speech increased, the identified rate was remarkably improved through loudness compensation. Moreover, the average recognition rate of the output speech signals after the loudness compensation based on the gammachirp filterbank was better than the system output based on the gammatone filterbank, which verifies the effectiveness of the method.

5. Conclusions

In this study, the gammachirp filterbank was applied to the multichannel compensation algorithm of digital hearing aids, and a new channel division method was proposed based on human hearing characteristics. The gammachirp filterbank was employed to decompose and synthesize the input signal, and personalized loudness compensation was performed on the speech signal in each subband, which could effectively supplement the speech energy loss in different frequency bands for hearing impaired patients. The new channel division method made it possible to flexibly adjust the subband filter bandwidth during the hearing aid fitting process, to complete an individual filterbank design, and to achieve signal reconstruction with fewer channels and to meet different audiogram matching requirements. Future studies that include extensive subjective experiments are warranted to optimize the performance of the proposed algorithm. The other possible research trend would be to focus on the data-driven channel division method and loudness compensation model.