Artifact Removal Method in Embedded Hearing Diagnosis System

1 Overview
Hearing and language are important means for humans to communicate with each other and understand the world, but the haze of ear diseases and hearing impairments plagues humans. For newborns, the earlier the hearing loss is diagnosed, the greater the chance of cure. Therefore, it is particularly important to use a hearing screening device to screen their hearing. Hearing diagnostic equipment can detect the extent of their hearing loss early and provide timely help. For adults and the elderly, good hearing diagnostic equipment can not only help them understand their own auditory system, but also assist hearing-impaired people in choosing appropriate hearing aids and help them resume a healthy life.
Otoacoustic Emissions (OAEs) is a type of audio energy generated in the cochlea and released into the external auditory canal through the ossicular chain and the tympanic membrane. It is one of the objective hearing test methods currently widely used in clinical practice. In the external auditory canal, since the intensity of the otoacoustic emission signal is very low, generally not exceeding 20dB SPL (where SPL means sound pressure level, which is the reference value used by national metrology departments to calibrate various hearing instruments), and the stimulus artifact is relatively much stronger, the suppression of stimulus artifacts has always been an important and difficult problem in OAEs measurement. This paper studies the OAEs detection algorithm and comprehensively applies the current signal processing methods in the embedded hearing diagnosis system, achieving good results in artifact elimination.

2 Composition of otoacoustic emission signals
The composition of evoked otoacoustic emission signals in the external auditory canal can be approximated by the model in Figure 1.

In Figure 1, G1 represents the transfer function of the acoustic signal through the external ear and middle ear, including the response of the external auditory canal, tympanic membrane and ossicular chain; G2 represents the reflection of sound energy in the auditory conduction pathway; L(t) represents the response signal in the external auditory canal after the stimulus sound propagates and reflects through the auditory pathway, i.e., the stimulus artifact; NL and NL(t) represent the cochlear response induced by the stimulus sound, i.e., the otoacoustic emission signal; N(t) represents the random noise in the external auditory canal, and R(t) represents the signal in the ear canal. R(t) can be expressed as:
R(t)=L(t)+NL(t)+N(t)
Since the intensity of otoacoustic emission signals is very low, generally not exceeding 20dB SPL, the stimulus artifact accounts for a large proportion of the induced otoacoustic emission signals. In OAEs measurement, the stimulus artifact must be effectively suppressed.

3 Artifact elimination scheme
Under normal hearing conditions, the delay time of TEOAEs (Transient-Evoked Otoacustic Emissions) relative to the stimulus signal is 3 to 5 ms, and the duration is about 15 ms. The frequency components of TEOAEs are closely related to the composition of the stimulus signal. The latency of TEOAEs is related to the frequency. In terms of time, the high-frequency component appears first and the low-frequency component appears later. According to a report by British scholar Kmep: the latency of 5,000 Hz TEOAEs is 4 ms, while the latency of 500 Hz signals is about 12 ms. This phenomenon is mainly because the traveling waves of different frequencies travel different distances on the basilar membrane. The intensity of the stimulus sound also has a certain effect on the waveform and amplitude of TEOAEs. A wider frequency range of OAEs can be obtained at high stimulus intensity. As the stimulus intensity decreases, the energy of OAEs is increasingly concentrated on several frequencies close to spontaneous otoacoustic emissions. Using a higher stimulation intensity can obtain more comprehensive information about the cochlea, but it also increases the intensity and duration of the stimulation artifact (direct reflection of the external auditory canal and middle auditory canal to the stimulation sound). Therefore, the stimulation artifact needs to be eliminated during TEOAEs testing. The correlation coefficient is shown in Figure 2.

3.1 Time Domain Windowing Method
The artifacts can basically disappear after about 5 ms from the start of stimulation, while TEOAEs have a latency of 3 to 5 ms. Since the artifact components are mainly present in the period 0 to 2.5 ms from the start of stimulation, the signal in this period is often set to 0; while the period 2.5 to 5.1 ms is the area where TEOAEs and artifacts coexist. As time increases, the artifact components gradually decrease while the TEOAE components gradually increase. In the period of 5.1 to 20 ms, TEOAEs are the main components, because this time domain window is selected as a cosine rise or fall rectangular window, and the cosine rise or fall time is generally 2.5 ms or 2.6 ms. The main advantage of this method is that it can remove the stimulus artifacts very cleanly, and the method is simple, but some short-latency components (generally high-frequency components) in TEOAEs are also removed at the same time.
3.2 Nonlinear Differential Average Method
Nonlinear differential average (DNRL) is also called "derived sublinear response". The basic principle is to express the acquisition wave x(t) as the superposition of the reflection wave R(t) and the otoacoustic emission wave OAEs(t):
x(t)=R(t)+OAEs(t)
Assume that the reflection wave R(t) is mainly a linear component, which increases in proportion to the stimulus intensity; while OAEs(t) is saturated within the appropriate stimulus intensity range, and it basically does not increase with the increase of stimulus intensity, that is, it has nonlinear characteristics. Therefore, the four adjacent stimulation records are taken as a group, in which the intensity and polarity of the last three stimulations are the same, and the intensity of the first stimulation is three times that of the last three stimulations and the polarity is opposite. Finally, the four records are accumulated and averaged, and then:

doubling the result is the desired OAEs(t) waveform. The outstanding advantage of the DNLR method is that when the stimulus intensity is large, that is, when TEOAEs show strong saturation nonlinearity, it can effectively remove artifacts. However, it also has the following two disadvantages:
① The actual TEOAEs have not only nonlinear components, but also linear components. When the stimulus intensity is high and TEOAEs are close to the saturation zone, the nonlinear component is dominant, and the method works better at this time; but when the stimulus intensity is low and TEOAEs are in the non-saturation zone, the linear component is dominant, and in this case the DNLR method is not feasible.
② Compared with the coherent averaging method using the same stimulus, the DNRL method will reduce the signal-to-noise ratio, and the amplitude of the TEOAEs obtained after processing will be reduced.

4 Experimental results
Test conditions: The test was carried out on the embedded hearing diagnosis system platform designed by this laboratory; the background noise was less than 50 dB SPL; the stimulus signal intensity was 80 dB SPL.
In the design of the signal processing scheme, the time domain windowing method and the nonlinear differential averaging method were used in combination: first, the cosine rise was 2.5 ms cosine rectangular window to filter out the artifacts of the latent period; then the nonlinear differential averaging method was used to filter out the artifacts after the latent period. In this way, the advantages of the above two methods of removing artifacts are combined, and their respective shortcomings are made up, a good signal-to-noise ratio is achieved, and the purpose of eliminating artifacts is achieved.
The transient evoked sound was tested on unilateral ears of 10 young people, and the signal-to-noise ratio was estimated. The results are listed in Table 1.

In Table 1, the data in the first row are the signal-to-noise ratios obtained by using a simple time-domain windowing method; the data in the second row adopt a comprehensive method, that is, the time-domain windowing method and the nonlinear differential averaging method are used successively. As can be seen from Table 1, the signal-to-noise ratio has been significantly improved.