6dB SNR improved 64 Channel Hearing Aid Development using CSR8675 Bluetooth Chip

Transcription

1 016 International Conference on Computational Science an Computational Intelligence 6B SNR improve 64 Channel Hearing Ai Development using CSR8675 Bluetooth Chip S. S. Jarng Dept. of Electronics Eng. Chosun University Gwangju, South Korea A. Amatya, Y. J. Kwon Hearing Impairment & Rehabilitation Res. Ins. Algorkorea Co. Lt. Seoul, South Korea D. S. Jarng Dept. of Computer Science Biola University Los Angeles, CA, USA Abstract Hearing loss affects hearing impairment at iversely ifferent frequencies, so it is crucial that a goo hearing ai shoul have high frequency resolution. We have evelope a 64 channel hearing ai which provies precise control of 15Hz resolution over 8000Hz frequency spectrum with nonlinear compression. Probabilistic noise reuction algorithm an feeback cancellation base on frequency omain analysis were evelope, which enhances speech clarity while minimizing unesirable ambient noises. The hearing ai can also be controlle an calibrate wirelessly via Bluetooth. Keywors-15Hz resolution; non-linear compression; auiogram; 64 channel spectrum; noise reuction; feeback cancellation; CSR 8675; Bluetooth I. INTRODUCTION Moern hearing ais are not only concerne with increasing the auibility of soun, but also the clarity of speech. Most commercially available hearing ais toay have a frequency resolution larger than 500Hz. In many cases a hearing impairment may have varying levels of hearing loss within a 500Hz range across the 8 khz spectrum. Applying the same amplification across the 500Hz woul be sufficient enough to make the soun auible, but woul not necessarily improve an might even egrae the clarity of speech. Therefore a higher resolution becomes a necessity in hearing ai practice. To attain high quality of speech, the 8000Hz frequency spectrum nee to be ivie into 64 channels with each channel emarcate by a 15 Hz step size. To improve the speech quality further, a probabilistic noise reuction algorithm was esigne to suppress ambient noises while preserving speech. A feeback cancellation algorithm base on the energy level of the input soun signal was use to suppress acoustic feeback. Furthermore, the wireless control of hearing ais greatly increases the usability of hearing ais. Hence the Bluetooth capability of CSR 8675 chip was utilize to a features to wirelessly control volume level, an enable or isable noise reuction an/or feeback cancellation, an perform the remote fitting of hearing ais. Hearing losses can be classifie into three main categories (mil, moerate, severe) base on type of loss[1]. It is necessary to know a patient s hearing level for precise fitting of his/her hearing ai. We esigne an In-Situ hearing test metho with an anroi app that measures hearing threshols at ifferent frequencies to get the auiogram, an the hearing ai fitting parameters are calculate base on this measurement. The in-situ hearing test provies an explicit auiometric measurement of a customer s hearing level. We use this measurement to prouce fitting curves that woul be suitable for the customer. II. 64 CHANNEL NON-LINEAR COMPRESSION Dynamic range of hearing, also referre to as resiual ynamic range, allues to the range between the maximum an minimum auible threshol levels of hearing. The most common type of hearing loss is the sensorineural hearing loss, which reuces the ynamic range. This means that softer souns are barely auible, while lou souns remain too lou []. It can cause conversational voice to become inauible or isconcerting. Merely amplifying the soun woul help soft soun auible, but lou soun might become painfully louer. Hence speech soun signal has to be compresse to maintain a balance between lou an soft souns so that they fall within the specific ynamic range of a hearing impairment. There are two istinct types of compression 1) Linear compression ) Non-linear compression In case of linear compression the maximum output is limite to a certain value an all output souns are less than or equal to that value, but with uniform gain applie. However, in non-linear compression, varying amplification /16 $ IEEE DOI /CSCI

2 is applie to ifferent levels of soun input. Fig. 1 shows the input/output curve as a non-linear compression example. Figure 1. Non-Linear compression of the hearing ai. In non-liner compression, for example, softer soun gets a higher amplification, while louer soun gets a lower amplification, so that all the possible input souns fit within the ynamic range of the user, as shown in Fig. There is still an upper limit for the louest possible output though, an all soun outputs are less than or equal to this value. Fig. shows an input-output (I/O) curve (upper) an an input-gain (I/G) curve (lower) for non-linear compression. Both the curves show how ifferent values of gain are applie to input souns of ifferent levels. The output is the input plus the gain. The compression region consists of four sections, where each section represents a range of input auio intensity. The segments labele s1, s, s3, an s4 in Fig are known as squelch, linear, non-liner, an saturation segment respectively. The squelch threshol (SQTH) represents the upper limit of the squelch region (s1), up to which input souns are suppresse to reuce unwante ambient noises. The linear region (s) is between SQTH an the lower threshol (LTH) point. Within this region a constant gain is applie to the input signal. In the next segment, the nonlinear region (s3) lies between the LTH an the upper threshol (UTH) point. Varying levels of amplification are applie to the input in this region. The UTH point represents the optimal amplification that can be applie to the lou input signal. A low level gain (LLG) controls the gain threshol within the linear region. Inputs beyon the UTH lie in the saturation region (s4) where the output is limite to a maximum level. By varying the high level gain (HLG), the saturation level can be set to any esire an fixe level. Base on the threshol levels at each frequency channel, the non-linear compression parameters are calculate for the entire frequency spectrum. A. Calibation The compression scheme escribe above is use to calculate varying gain values for all the frequency channels in the hearing ai. The avantage in our hearing ai firmware esign is the high frequency resolution, provie by iviing the 8000Hz, half of the 16000Hz sampling frequency, frequency range into 64 channels. Each channel has a frequency resolution of 15 Hz. In orer to calculate the proper non-linear compression gain parameters we nee accurate measurements of patient s hearing threshols. This can be one either by entering the patient s auiogram information, or by performing an In-Situ self-hearing test. This can be performe by sening appropriate commans to the hearing ai through Bluetooth communication. The anroi app monitor interface use for performing the In- Situ hearing test is shown in Fig.3. Figure. Four ifferent compression regions. Figure 3. In-Situ Hearing Test. When a user touches a point on the auiogram graph of the smartphone screen, a sine wave tone of corresponing 144

3 frequency inicate on the x-axis at the bottom, an intensity inicate on the y-axis at the left, is playe from the hearing ai. The comman ata to play this tone is passe wirelessly from smartphone to hearing ai via Bluetooth. The user nees to fin the hearing threshol of at least four frequencies, 500Hz, 1kHz, khz an 4kHz. The app will then calculate the optimal fitting parameters on the basis of this hearing threshol ata. Couple with the patient s hearing ability, both the microphone an the receiver use in the hearing ai are calibrate at the same 15 Hz precision, obtaining an accurate representation of the capabilities of the microphone an the receiver. With this calibration ata an an accurate auiogram, we can calculate the non-linear compression gain parameters require to prouce input-gain curves as shown in Fig using the formula; Gain = S * X+B. (1) where, X = x n + TM S =.0 * s n * C. B = b *.0 * 40. C = Calibration constant which is a combination of microphone an receiver sensitivity values b = (y s n * X) * C. B. Bluetooth Control The CSR 8675 chip s Bluetooth capabilities have been use to control the hearing ai wirelessly. Firstly the hearing ai nees to be paire with the smartphone like any normal Bluetooth evice. Then our specialize anroi app can be use to sen commans to the hearing ai wirelessly. An example of the anroi app s interface for controlling the hearing ai is shown in Fig. 4. saves fitting ata to hearing ai flash memory (EEPROM) respectively. Other hearing ai control buttons inclue Normal, which turns off noise reuction (NR) an feeback cancellation (FBC), the NR On an FBC On buttons which turn on noise reuction an feeback cancellation respectively, an the NR an FBC On button, which turns on both noise reuction an feeback cancellation. The Volume Up an Down buttons are use to wirelessly increase or ecrease the volume respectively on the hearing ai. These are some of the main components of the interface use for Bluetooth control. III. NOISE REDUCTION High frequency resolution non-linear compression alone is not sufficient to increase the clarity of speech for hearing impairments. People with normal hearing ability are naturally filtering out environmental noise an isturbances, while people with hearing loss are not as capable of oing the same. Hence a noise reuction algorithm plays a significant part in filtering out unwante ambient noises. We take the avantage of the 64 channels present in our hearing ai to esign a probabilistic noise reuction algorithm [3]. In this algorithm, we continuously etermine the minimum power of input signals an calculate a smoothing parameter base on this which helps us get the noise estimate. The noise estimate is then use to calculate the Signal to Noise Ratio (SNR), which etermines the presence or absence of speech. A lower gain is selecte if the input is etermine to be noise, an a higher gain is selecte if it is etermine to be speech. The output has minimal istortions an is clear enough for practical purposes. A. Backgroun Consier a noisy speech signal, y(n), y(n) = x(n) + (n). () Where x(n) an (n) enote iscrete time signals of clean speech an noise respectively. The noisy speech signal, y(n), is split into overlapping frames using a winow function. A Short-Time Fourier Transform (STFT) written as, N 1 n0 j( ) kn N Y y( n lm) * h( n) * e. (3) is use to analyze the overlapping frames, where k = frequency bin inex, l = time frame inex h = Analysis winow of size N, m = framing step (Number of samples separating two successive frames) Figure 4. Bluetooth control interface. The READ comman button reas the fitting ata from hearing ai memory to the anroi app. The WRITE or BURN button writes fitting ata to hearing ai RAM or Analyzing () using the STFT equation given in (3), we have Y( = X( + D(. (4) 145

4 B. Noise Reuction Algorithm Equation (4) shows that a noisy signal is the combination of the esire speech signal an environmental noise. We assume that the noise environment varies slowly compare to speech, an that the noise spectrum remains practically stationary uring speech activity. It is also assume that the intensity of the noise signal has the behavior of Gaussian istribution in each frequency ban, an that the noise uncorrelate with the speech signal. An estimate of the STFT of clean speech shown in (4) is obtaine by applying a specific gain function to each spectral component of the noisy speech signal [4], an is given by, Xˆ G( * Y. (5) where, k = frequency bin inex l = time frame inex Y = STFT of the noisy signal Xˆ = optimal spectral amplitue of the voice signal For Y( that are etecte as speech the gain is high, while for those values etecte as noise, the gain is low, hence causing attenuation of noise. The task then becomes etermination of whether Y( is noise or not, an etermination of the gain. We start by recursively smoothing of the input noisy signal S( = s *S( l-1)+(1- s )*S f (. (6) where, S = smoothe energy value of input signal s = smoothing parameter S f = current unsmoothe energy value of input signal, Y( The minimum of the local energy S min ( shown in Figure 5 is tracke as follows: Figure 5. An example of the noisy speech spectrum an the minimum local energy. The ratio between Y( an its erive minimum of the local energy is enote as, S r ( = Y( / S min (. (11) The conitional speech presence probability p ˆ is estimate as pˆ pˆ ( l 1) (1 ) I(. (1) p where I( =1, S r ( > an I( = 0. p is smoothing parameter set at 0. an is the threshol value for S r set at 5. The noise spectrum ˆ is recursively estimate while a smoothing parameter is upate with the conitional speech presence probability. ˆ (1 ) pˆ (. (13) s ˆ s l 1) [1 s ] Y(. (14) p S min ( = min{s min ( l-1), Y( }. S tmp ( = min{s tmp ( l-1), Y( }. If mo( = 0, then S tmp ( is re-initialize. S min ( = min{s tmp (, Y( }. S tmp ( = Y(. (7) (8) (9) (10) If the spectrum of the input noisy speech is higher than the estimate noise spectrum at a particular frequency ban, we assume speech to be present in this ban an apply higher gain. We then calculate the transient beam to reference ratio as S(. (15) Y ˆ We then calculate the likelihoo of speech presence by calculating ( = 1 for Y > 0 an = 0 otherwise, then taking average, as k1 k k 0. (16) k0 k

5 where k0 an k1 have been chosen to be 3 an 55 between 1 an 64 respectively. To calculate speech absence probability we first calculate s as Y. (17) s ˆ For where 678 TH TH 0. the apriori speech absence probability is calculate as 1 s ˆ(, ) max,1. (18) q k l 1 1 where 1 = To calculate the speech presence probability which is the focal point of the noise reuction algorithm, we first calculate the aposteriori SNR as l l 1) [1 ] Y( ). (19) The smoothing parameter is calculate as (1 ) * p(. (0) where = an p( is the speech presence probability which is calculate as qˆ( p( 1 (1 1 qˆ( 1 ( k, * e. (1) Figure 6. An example of the noisy speech spectrum an the enhance speech spectrum. C. Noise Reuction Results The noise reuction algorithm was at first simulate in Matlab an then implemente in the CSR 8675 chip. An anroi app was use to generate white noise, an a single syllable a soun was mae as the speech input. The output from the CSR 8675 chip fe into a measuring microphone through a stanar -cc coupler an the microphone output was observe in the oscilloscope. Observing peak-to-peak values for noise an speech, the SNR for noise reuction off was seen to be 7.3 B an for noise reuction on it was seen to be 15.4 B, giving an improvement of 8.1 B. Figure 7 shows the oscilloscope result with noise reuction turne off an Figure 8 shows it with the noise reuction turne on. * (. () 1 Y. (3) GH1( l 1) l 1) (1 ) max{ 1,0}. (4) Figure 7. Oscilloscope output with noise reuction turne off. where = G H1 t 1 e exp t. (5) 1 (, ) k l t G( p(, ) 1 (, { 1 (, )} k l p G k H k l G. (6) min where G min = G( is the gain that gets applie to each frequency channel of the l th input frame. Figure 6 shows a plot of noisy an enhance speech. Figure 8. Oscilloscope output with noise reuction turne on. 147

6 IV. FEEDBACK CANCELLATION Output souns from the hearing ai receiver which fin their way back into the microphone can prouce a high pitche tone referre to as feeback. For example, if the hearing ai is a not a customize ear mol, but a generic esign, then subtle facial movements can cause the hearing ai to unsettle from its ieal position causing feeback to occur. If the fitting for the hearing ai is poorly one, then feeback can occur as well. Feeback can be reuce by physical improvements to the hearing ai, such as esigning the shell to fit snugly into the user s ear canal, using a lower gain, or esigning the vent properly. However, it is possible to take avantage of a higher number of frequency channels in a igital hearing ai to implement feeback cancellation feature only by firmware. Therefore a feeback cancellation algorithm was implemente to avoi this untowar isturbance. A. Feeback Cancellation Design The energy (square magnitue) of the complex (frequency omain) ata of the input signal is use to etermine if there is feeback present in the signal or not. E = real(x) + imag(x). (7) The minimum of the local energy inicates the presence of feeback within the system. A series of tests for feeback presence in the hearing ai were performe an the energy present within the signal was monitore. Figure 9 shows istribution of energy in the input signal with respect to frequency channel for various types of input auio, with an without feeback. The feeback suppression algorithm starts by splitting the frequency spectrum into five istinct regions. The energy level within each region is checke. A threshol value for each region was empirically etermine. If the energy value in a region went above this threshol, it was etermine to be ue to feebac an hence an attenuation is applie. When the feeback soun isappears the energy value in each region comes below the threshol for that region, an then the attenuation is remove. V. CONCLUSIONS A 64 channel hearing ai with non-linear compression provies a high level of speech clarity than other types of hearing ais available ue to its 15 Hz frequency precision. The clarity arises ue to a combination of the compression with the backgroun esign setup which also functions on a 15 Hz precision. The speech clarity can be further improve by aing a probabilistic noise reuction algorithm [5,6] an a feeback cancellation algorithm. The implementation of these algorithms an their effectiveness is augmente by the higher level of frequency precision. The availability of wireless control via Bluetooth makes the hearing ai more user frienly an appealing the general populace. Figure 9. 1-Strong Feeback; -Weak Feeback; 3-No feeback; 4- Feeback in Ear; 5-Voice only; 6-Voice with feeback. ACKNOWLEDGMENT This stuy was supporte by research fun from the ministry of commerce, inustry an energy (MOCIE Korea) 015 core meical evice commercialization technology evelopment project (smartphone controlle 64 channel igital hearing ai with 6B voice SNR (Signal to Noise Ratio) improvement: project number ) in 015. REFERENCES [1] R. Sanlin, Hearing Instrument Science an Fitting practies, n e., International Hearing Society, 1996, pp [] S. Banerjee, The Compression Hanboo 3r e., Starkey Hearing Research & Technology. [3] S. S. Jarng, C. Swanson, F. Lee an J. Zou, 64 Bans Hybri Noise Reuction an Feeback Cancellation Algorithm for Hearing Ai, International Journal of Control an Automation, vol. 7, No.1, pp , 014. [4] I. Cohen an B. Berugo, Speech enhancement for non-stationary noise environments, Signal Processing, vol. 81, no. 11, pp , 001. [5] Y. Ephraim an D. Malah, Speech enhancement using a minimummean square error short-time spectral amplitue estimator, IEEE Trans Acoustics Speech an Signal Processing, vol. 3, no. 6, pp [6] S. Kamath an P. Loizou, A multi-ban spectral subtraction metho for enhancing speech corrupte by colore noise, Proceeings of ICASSP-00, Orlano, FL,