Journal of the International Telemedicine Academy

Contactless Hearing Aid for Infants Employing Signal Processing Algorithms

Maciej KULESZA¹, Piotr DALKA¹, Bozena KOSTEK^1,2

¹ Multimedia Systems Department, Gdansk University of Technology, Narutowicza 11/12, 80-952 Gdansk, Poland

² Institute of Physiology and Pathology of Hearing, Warsaw, Poland

{maciej_k, dalken, bozenka}@sound.eti.pg.gda.pl

Abstract:

The proposed contactless hearing aid is designated to be attached to the infant's crib for sound amplification in a free field. It consists of 4 electret microphone matrix, and a prototype DSP board. The compressed speech is transmitted and amplified via miniature loudspeakers. Algorithms that are worked out deal with parasitic feedback, which occurs due to the small distance between microphone and monitors and potentially high amplification required. The beamforming algorithm is based on an artificial neural network (ANN). The ANN is used as a nonlinear filter in the frequency domain. Principles of algorithms engineered and the prototype DSP unit design are presented in the paper. Also, results of experiments simulating the real-life conditions are analyzed and discussed.

---

Introduction

An early intervention in rehabilitating an infant having a hearing loss is of a great importance. However this poses a unique set of problems. Since it is not possible to evaluate infant's hearing loss by subjective methods, and then to check the validity of the prescribed gain, thus evoked potential (ABR - Auditory Brainstem Response) assessment is required to measure the hearing loss and establish hearing aid fitting targets [1]. An audiogram is to be predicted on the basis of the ABR assessment, and then the appropriate amplification of the hearing aid is set up. This is especially important in terms of the infant's speech skill development and understanding. It should however be remembered that hearing loss evaluation and fitting are only first elements of the rehabilitation chain. Prescribing a typical hearing aid for infants is not very practical. An infant tries to take the hearing aid out, plays with it, etc., which may cause changes to the hearing aid settings or may even damage the device. The size of the behind the ear (BTE) hearing aid is quite large in comparison to the infant's head/pinna, and it is uncomfortable in case when an infant wants to lay its head on a side. On the other hand, any in-the-ear (ITE) or insert (ITC - in the canal or CIC - completely in canal) hearing aids are not recommended to wear for an infant because of the growing of the ear canal and changing its anatomical shape in time. In addition, a hearing aid may cause some malformation of the bony ear canal.

Due to the rapid development and an increase of power of miniature signal processors along with DSP algorithms it is possible to think up a totally different approach to the hearing aid for infants. The proposed contactless hearing aid is designated to be attached to the infant's crib for sound amplification in an acoustical field. It transmits an amplified and compressed signal of an infant mother's speech (or any person taking care of an infant) via miniature loudspeakers. A design of the dedicated extension card for the TMS320VC5509A DSP development board has been made at the stage of research work preparation. The prototype was engineered at the Multimedia Systems Department, to implement and test algorithms enabling functioning of the contactless hearing aid [2].

Algorithms that are worked out deal with some obvious limitations of the free-field hearing aid such as parasitic feedback, which occurs due to the small distance between microphone and monitors and potentially high amplification needed. That is why one of the algorithms implemented is beamforming which controls a feedback between microphones and monitors. For this purpose a matrix of 2 electret microphones has been employed. The beamforming algorithm is based on an artificial neural network (ANN), thus the main problem concerns choosing appropriate feature vectors that are feeding the given algorithm inputs. The ANN is used as a nonlinear filter in the frequency domain. The main task of the spatial filter is to estimate the desired signal arriving from the front direction. It is neither desirable nor possible to completely attenuate signals from lateral and backward directions. Because spatial filter works in the frequency domain, it is assumed that each spectral component representing signals coming from unwanted directions is to be attenuated by at least 40 dB [3][4][5]. The spectral components representing signals coming from the forward direction should remain unaltered. The effectiveness of the algorithm engineered and the resultant speech intelligibility depends on the proper decision made by the neural network, thus the learning procedure of the ANN is very important. This decision is made basing on the values of the parameters of sound that are similar to those used by the human auditory system. These parameters represent both interaural intensity ratio and interaural time difference. In the testing phase various combinations of signals are introduced to the ANN inputs and the algorithm is checked as to its effectiveness.

Apart from beamforming, various techniques based on signal processing are commonly employed in digital hearing aids in order to prevent the occurrence of the acoustic feedback [6][7]. Two such techniques (adaptive notch filtering and adaptive feedback cancellation) were studied and the former one was chosen for experiments and implementation as an additional protection against feedbacks.

Another algorithm implemented in the hearing device is the voice activity detector (VAD). Its goal is to prevent the amplification of loud unexpected sounds as well as any other acoustic signals that should not be presented to the infant.

This paper is structured as follows. In Section 2, a brief description of the contactless hearing aid set-up is introduced. In Section 3, the beamforming algorithm based on ANN is described. Other methods for the acoustic feedback elimination are discussed and compared in Section 4. The Section also contains preliminary experimental results. Section 5 presents practical implementation remarks, especially concerning the design of an extension card for the digital signal processor. Finally, in Section 6 the authors summarize the results obtained and outline future research related to the contactless hearing aid (PCT patent pending).

Contactless Hearing Aid

In contrast to all standard digital hearing aid solutions, where a microphone, digital signal processor and receiver are enclosed in one shell, the contactless hearing aid set up comprises of the following three separated modules:

• microphone array (4 electret microphones) with preamplifiers mounted in front of the infant's bed,
• DSP unit responsible for signal processing,
• miniature loudspeakers mounted near the infant's head.

In Figure 1, the infant's crib equipped with a contactless hearing aid is presented.

Figure 1. Contactless hearing aid application

The distributed structure of the device allows not only for its convenient installation in the infant's crib but is also essential for reducing the possibility of the parasitic feedback occurrence, as the microphones and loudspeakers can be easily separated from each other. Moreover, the DSP unit may be mounted far from the infant's body or even be hidden under the bed in order to minimize its eventual negative influence on crib ergonomics.

Since the hearing aid is not a simple sound reinforcement system, advanced signal processing techniques have to be applied in order to meet specific requirements of the signal capture and hearing loss compensation. The functional diagram illustrating the processing modules of the hearing aid is shown in Figure 2.

Figure 2. Block diagram of signal processing in the contactless hearing aid

The first module comprises of four omni directional microphones which capture sounds coming from all directions. In order to attenuate signals emitted through the loudspeakers as well as the signals emitted by the infant itself or any other undesired sources, the spatial filtration module is employed. Spatial filtration (beamforming) takes the advantage of the fact that the distance from the source to each microphone in the array is different, which means that the signals captured by the microphones will be phase-shifted replicas of each other. Knowing the amount of the phase-shift at each microphone in the array is sufficient to calculate the direction [8]. It is assumed that the desired signal (e.g. coming from a speaker) that should be presented to the infant's ears may come only from the sources located in front of the crib. Thus the adaptive spatial filtration algorithm tracks the speaker only in a limited area. As the basic aim of this technique is to attenuate all signals that are undesired for further processing, this can also be viewed as a particular method for the parasitic feedback cancellation.

In order to prevent the amplification of loud unexpected sounds as well as any other acoustic signals that should not be presented to the infant, the voice activity detector (VAD) is employed. This algorithm operates in the frequency domain and takes the properties of the captured signal spectrum into account to decide whether the signal should be further processed [9]. Furthermore, based on the detector state particular groups of electronic components (e.g. power amps, digital to analog converters) are either activated or disabled. It allows minimizing the power consumption of the device, which is especially important when the hearing aid is battery powered during outdoor operation.

Despite the beamforming technique, it can be expected that in some particular situations the acoustic feedback may occur. Thus the captured signal is continuously analyzed in order to eliminate this parasitic phenomenon. Although the signal is further processed by two separated modules for left and right channel respectively, and then emitted through the miniature loudspeakers, the acoustic feedback elimination procedure is applied to the mono signal just before the separated channel processing occur. This seems reasonable since the characteristic of the feedback is common for these two channels.

The last module in Figure 2 implements signal processing algorithms that are typically employed in digital hearing aids. In the first step the signal is divided into subbands. Further, either constant amplification or dynamic processing with compression characteristic is applied in each signal subband based on the characteristics of an infant's hearing impairment. One can notice from the Figure 2 that hearing loss is compensated for each ear independently. Finally, two processed signals are emitted through the miniature loudspeakers mounted near the infant's head.

Neural Beamformer

Automatic identification of sound sources direction is still an unsolved problem in many real-life applications, such as for example, hearing prostheses or contemporary teleconferencing systems. The main reason for this is background noise, high reverberation and/or with many concurrent speakers. One approach to reducing this noise is to provide directional field of hearing. Source identification (spatial filtration) system should allow for tracking a target speaker automatically without much delay in order to avoid picking up concurrent speakers by the same microphone channel. This may be done in various ways, however, generally two approaches can be found in literature. One of them is a conventional approach to this problem based on delay summation algorithms, superdirective arrays and adaptive algorithms, non-linear frequency domain microphone array beamformers, etc. [10][11] [12][13][14][15][16]. The effectiveness of these algorithms decreases however while performing in reverberant environments. Examples of such algorithms were broadly reviewed in literature, thus they will not be recalled here. The second approach to this problem was proposed in the Multimedia Systems Department, GUT in collaboration with the Institute of Physiology and Pathology of Hearing, Warsaw in previous studies. Namely Artificial Neural Networks (ANNs) have been applied for the purpose of the automatic sound source localization [3][4][17][18][19][20][21][22][23]. Since the current study requires an effective spatial filtration algorithm, thus the approach based on beamforming employing ANNs was reviewed here within the context of contactless hearing aid. The ANN was used as a nonlinear filter in the frequency domain (also time domain neural beamformer is easy to implement).

The first step of experiments consisted in extracting feature vectors to be fed to the ANNs. During the feature extraction process the signal acquired was divided into frames of the length of 256, 512 or 1024 samples. The feature set was based on previously defined parameters under the assumption that a neural network provides an effective non-linear filtering algorithm of an acoustic signal transformed into the frequency-domain [3][4][5][22]. It was assumed that the number of microphone channels should be limited to two. Signal arriving at both microphones can be written as:

; (1)

where:

• l(t), r(t) - signals received by the left and right microphones,
• s(t) - desired signal arriving from the front direction,
• n_l(t), n_r(t) - signals coming from the lateral or backward directions arriving to the left microphone and to the right microphone. These signals are treated as noise.

The main task of the spatial filter is to estimate the desired signal s(t) arriving from the forward direction. Because spatial filter works in the frequency domain, it is assumed that each spectral component, which represents signals coming from unwanted directions should be attenuated by at least 40 dB (see Figure 3). In Figure 3 a prototype spatial characteristics is shown. The spectral components that represent signals coming from the forward direction should remain unchanged. This can be described by the following expressions:

(2)

where:

• i - spectral component index,
• , - estimates of a signal component Lⁱ in the left, and Rⁱ in the right channel,
• g(i) - attenuation coefficient of noisy components described by the following formula:

(3)

Figure 3. Desired directional characteristic (the same for all frequencies). x - axis represents angle, y - axis represents attenuation in [dB]

The effectiveness of this algorithm and the resultant speech intelligibility will depend on the proper decision made by the neural network, so the learning procedure is very important. This decision is made basing on the values of some parameters of sound that are similar to those used by the human auditory system. These parameters represent both interaural intensity ratio and interaural time difference. The first parameter, which expresses the interaural spectral magnitude ratio, is described by the following expression:

(4)

where:

• i - spectral component index,
• Lⁱ, Rⁱ - left and right signals for the ith spectral component,
• Mⁱ -magnitude ratio for the ith spectral component

The second parameter, which expresses the interaural phase difference is described by the following expression:

(5)

where:

• - denotes the signal phase,
• Aⁱ - phase difference of the ith frequency component of left and right channels

The third parameter used in learning phase is defined as:

(6)

where: Dⁱ - relative ratio of the ith spectral component for the left and for the right channel.

It can be shown that parameters described by Eqs. (4) and (6) are in a simple functional relationship and therefore one of them is superfluous. In such a case, parameters representing a single spectral bin can consist of parameters given by Eqs. (4) and (5).

During the learning phase the Mean Square Error (MSE) was observed. As seen from Eq. 7 MSE represents the squared error between the current value at the output of the network o and the desired response of the network d.

(7)

where P is the number of training patterns, and K denotes the number of outputs.

Feed-forward ANNs

The proposed neural network structure and its properties were such as follows: one hidden layer consisted of 9 neurons, the standard error backpropagation algorithm with momentum was used in the learning phase. The BP learning algorithm parameters were as follows: h = 0.5 (learning rate); a = 0.01(momentum ratio). Spectral components were obtained with 512 point FFT procedure using Blackmann window with an overlap of 256 samples. The training file consists of logatoms of every 15⁰ elevation. Each direction was represented by 10 sound examples (5 female and 5 male voices). In addition sounds from ą5° were used in this phase. These directions were treated similarly to 0° direction, thus the gain factor was equal to 1, whereas for other directions a value of 0.01 was used.

In the testing phase various combinations of signals were introduced to the ANN inputs. Namely such signals as: tones, tone plus noise, a phoneme (logatom) plus tone, a phoneme plus noise, phonemes and phrases were employed as testing material. Always one of the signals was coming from the front direction (0⁰), and the other was the unwanted one and was localized at the angle between 15⁰ to 90⁰ (horizontal plane). An example of spatial characteristics obtained after the learning phase are presented in Figure 4. As expected sharper minima and maxima were obtained for higher frequency spatial characteristics for the whole angle range. The slope of low frequency characteristics for 15⁰-90⁰ azimuth is very smooth.

Figure 4. Spatial characteristics of the ANN based filtration algorithm obtained with a multi-tone signal

In Figure 5 an example of a signal spectral representation (sonograms) before and after processing are shown. In Figure 5a combination of signals that was processed by the neural beamformer is shown. In this case the target signal was a logatom and the disturbing one was a 250 Hz harmonic tone. As seen from Figure 5 the disturbing signal is eliminated, but the proposed algorithm causes some distortions that are noticeable in the spectral domain. As seen from the sonogram analysis the target signal has got a formant around the same frequency as such of the concurrent signal. That is why the algorithm after processing cuts off this frequency along with the formant. However the signal-to-noise ratio equals to -60dB, so the distortions do not influence substantially the overall audio quality.

a) b)

Figure 5. Spectral representation of signals (phoneme, 0°)+(signal f₀=250Hz, azimuth 45°), before processing (a), after processing (b)

After processing various combinations of signals and azimuths it was observed that worse filtration effects were observed when a concurrent signal was close to the target signal (15° azimuth). In this case the dependence of the filtration effects on the character of the signal was also noticed. It can be also observed that definitions of parameters (Eq. 4) and (Eq. 5) cause that signals of the same spectrum composition coming from concurrent directions may not be effectively filtered out by such a beamformer algorithm. This is the most important drawback of the proposed method of spatial filtering, however in such a case a conventional beamformer does not perform well, either.

Acoustic Feedback Elimination Methods

This section describes two methods for acoustic feedback elimination (adaptive notch filtering and adaptive feedback cancellation) based on direct signal processing. Such a method forms an additional protection against feedbacks in the designed contactless hearing aid. Adaptive feedback cancellation turns out to be more suitable for DSP implementation, thus preliminary experiments regarding this method are also presented in this Section.

Algorithms for feedback elimination

Standard methods of dealing with feedbacks based on direct signal processing (e.g. passband equalizing) are static and unable to adapt to changes occurring in the system itself (e.g. microphones and loudspeakers movements) or in the acoustic environment. Two dynamic methods are often used to limit feedbacks: adaptive feedback cancellation and adaptive notch filtering.

The adaptive feedback cancellation method is very similar to algorithms used in acoustic echo cancellation for teleconferencing systems. The idea is to accurately model the loudspeaker to microphone transfer function F and then use this model to remove all of the audio sent out the loudspeaker from the microphone signal. An illustration of the method is presented in Figure 6.

There are many methods available for estimating the coefficients of an adaptive filter F', for example NLMS (Normalized Least Mean Squares), RLS (Recursive Least Squares) [24]. However, the resulting estimators are biased because the source signal v and the loudspeaker signal u are correlated. The bias can be eliminated by reducing the correlation. This can be achieved directly in the signal loop (by delaying or non-linearly distorting the loudspeaker signal u) or in an additional identification loop (by means of prefiltering the input signal y and the output signal u which assures that both the source and the loudspeaker signals v and u are whitened) [7]. The latter variant is computationally more complex.

Figure 6. Diagram of the adaptive feedback cancellation method

Adaptive feedback cancellation requires a significantly more powerful digital signal processor than adaptive notch filtering. It is capable to eliminate any audible signs of a feedback at the cost of some minor sound distortions.

The goal of notch filters deployed in the electroacoustic forward path between the microphone and the speaker is to eliminate frequency components resulting from the acoustic feedback. In this method, local maxima of the signal amplitude spectrum are detected and classified whether they represent feedback components. If a maximum representing a feedback component is identified, a notch is deployed with a centre frequency equal to the frequency of the local maximum [25].

There are two main steps in automatic notching algorithms: feedback discrimination and notch deployment. Feedback discrimination is based on a few properties of feedbacks that are very useful in separating them from natural sound features. The amplitude of a feedback component rises monotically and exponentially, while its frequency remains constant, which is illustrated in Figure 7. There are usually no harmonics of a feedback component, however non-linearity of electroacoustic devices working with high-level signals can be responsible for creating them. After a notch is placed on a potential feedback, its amplitude not only decreases by some value but continues to decrease at an exponential rate. This helps to verify the correctness of the feedback discrimination [26].

Notch deployment algorithm determines the parameters of new notch filters and rules for their deployment. A notch filter cannot be too narrow because of gradual changes in the acoustic environment and because of a limited precision of frequency identification. The width of the filter should be equal to 5 or 10 Hz, which guarantees the high effectiveness of the feedback suppression for a longer period of time. The depth of the filter equal to approx. 6 dB is sufficient to bring a feedback frequency back into stability; deeper filters would only decrease the sound quality.

The amount of notch filters is usually limited to the range from 12 to 20, since this number is sufficient to suppress all feedbacks that occur simultaneously. Because of the gradually changing acoustic environment, frequencies of feedback components are never static. Thus already-allocated notch filters is redeployed if required. When a new feedback component is detected, it is checked whether a filter has already been deployed at such frequency. If yes, the existing filter is deepened. If there is a filter with the frequency similar to the frequency of the feedback, the filter is widened to cover both frequencies. If all filters are allocated then the oldest filter is reset and redeployed at the new frequency.

The computational complexity of the adaptive notch filtering method is rather low while its effectiveness in the feedback elimination is high. Furthermore, sound distortions introduced by this method are insignificant. The only disadvantage is the fact that in order to be detected and eliminated the feedback components must first appear in an audio signal and may be audible for a short time.

Because of very strict power consumption constraints placed upon the hardware used in the embedded hearing aid device and because of its computational capabilities, the adaptive notch filtering method is chosen for experiments and implementation.

Experiments

The experiments regarding acoustic feedback elimination were focused on the implementation of the feedback component discrimination algorithm on the PC platform. During the experiments, pieces containing both speech and music parts were played through a computer speaker and recorded with a microphone. The placement of the speaker and microphone and their output and input gains were altered in order to produce a large variety of feedbacks. The recorded sequences were then processed to identify all local maxima in the signal spectrum and to recognize any feedback components.

For the purpose of the acoustic feedback cancellation feedbacks can be divided into three groups, as illustrated in Figure 7. The first group contains feedback components with amplitudes rising slower than approx. 6 dB per one time frame of the signal (the frame length of 46 ms was used). Such components are considered to be parasitic if their amplitudes continue to increase for 6 time frames in a row. The second group is formed by the potential feedbacks with amplitudes rising faster than approx. 6 dB per time frame. These components are suppressed if their amplitudes rise monotically for 3 frames in a row. The last group consists of feedbacks which are impossible to track directly because their amplitudes increase from the background level to the maximum level allowed in the system almost instantly (during the length of one or two time frames). There is only one way of dealing witch such feedbacks: all components with amplitudes higher than a given threshold are unconditionally considered as feedbacks. All components classified as resulting from feedback are eliminated by a notch filter. The algorithm guarantees that the louder a component and the faster it rises, the shorter time is required to detect and eliminate it.

The experiments carried out show that the algorithm engineered is able to detect and identify feedback components with a good accuracy. In a test sequence lasting 1.5 min that was infected with feedbacks, the algorithm detected 37 frequencies on which feedbacks occurred. In order to determine the effectiveness of classification, the algorithm was used to detect feedbacks in the original sequence (without feedbacks). As a result, only 3 false detections were identified. The amount of false positives can be further reduced through the observation of a new component for a few frames after notch filter deployment. If its amplitude decreases at an exponential rate then the classification is correct.

Figure 7. Examples of slow, fast and instant feedbacks

Implementation

This Section presents algorithm implementation details and describes the dedicated extension card for the Texas Instruments DSP development kit.

Hardware requirements

Since the contactless hearing aid is an autonomous device all algorithms responsible for signal processing must be implemented in the digital signal processor (DSP). In order to select the appropriate processor architecture it is necessary to declare the dynamics of the analog to digital and digital to analog conversion. Concerning the dynamic requirements for speech signal processing it was decided that the 16-bit precision is sufficient. In addition it was assumed that the device should be operable even if the power line is unavailable for some time (e.g. outdoor). In this case the power consumption of the DSP is the next issue that must be taken into account since it directly influences the battery life. Another important requirement is connected with processor peripherals. As stated in Section 2, the contactless hearing aid employs four microphones (and two miniature loudspeakers) in order to provide phase shifted signals into the spatial filtration module. Therefore, DSP must provide appropriate peripherals allowing for combining at least four A/D and D/A converters.

It is worth mentioning that the engineered algorithms described in the paper were evaluated in the Matlab simulator employing floating-point numbers representation. Thus, it seems at first that it would be reasonable to select the floating-point DSP architecture as well, since it might reduce the implementation effort. Unfortunately, the power consumption of the floating-point DSPs is significantly higher than the fixedpoint ones and thus floatingpoint DSPs are not recommended for battery powered applications. Therefore, the 16-bit fixed-point architecture of DSP was found suitable for implementing contactless hearing aid algorithms. It has to be pointed out however, that in the case of the fixedpoint DSP implementation the special care must be taken in order to maintain the robustness of the algorithms originally designed for floating-point processing.

During the selection of an appropriate DSP for the contactless hearing aid application various families of processors manufactured by Analog Devices (ADSP218x, ADSP219x), Texas Instruments (TMS320C54xx, TMS320C55xx) and Freescale Semiconductors (DSP56300) were taken into account [27][28][29]. Based on the requirements described above the Texas Instruments TMS320C55xx DSP family was chosen as the most suitable for the contacless hearing aid application. It has to be mentioned that the precise computational complexity of the algorithms incorporated into the contactless hearing aid is difficult to estimate before the implementation happens. That is why the possibility of altering the clock rate (108/144/200 MHz) determining the performance of TMS320C55xx is one of its advantages. Furthermore these processors provide also an advanced power management allowing extending the battery life. It was finally decided that all algorithms are going to be implemented using TMS320VC5509A processor.

Extension card design

In order to allow implementing and evaluating the algorithms the TMS320VC5509A development board manufactured by Spectrum Digital was employed [30]. It has to be pointed out that this system provides only one stereo audio codec, and thus it is impossible to evaluate the ANN spatial filtration algorithm. Therefore a dedicated extension card that extends the basic functionality of the development board was designed. The architecture of the development system consisting of Code Composer Studio environment, DSP development board and designed extension card is shown in Figure 8.

Figure 8: Block diagram of the contactless hearing aid development system

The main purpose of the extension card is enabling the A/D and D/A conversion of four autonomous signals and providing them to the digital signal processor. One of the extension card modules consists of the set of electret microphones and line levels preamplifiers. Although microphones can be plugged directly through the minijack ports (e.g. popular computer microphones), the dedicated port for microphone array module is also available. In addition, it is possible to capture the line level signals for evaluation purposes in every channel. The A/D conversion is obtained using two stereo audio codecs (PCM3008) that transmit and receive 16bit audio samples over the I²S serial interface [31]. Moreover, the sampling rate may be set to the one of the following values: 8, 16, 24, 32, and 48 kHz. Because the codecs can operate only in the slave mode, all necessary clock signals are generated locally. It is seen in Figure 8 that transmitted signals are buffered in order to prevent signal degradation. The physical connection between the extension card and the development board is accomplished using dedicated 80-pins peripheral slot [30].

Although TMS320VC5509A processor has three McBSP (Multichannel Buffered Serial Ports), only two of them are involved during the operation. The DMA (Direct Memory Access) controller of the processor is responsible for handling samples and feeding them to the appropriate buffers [32]. Then samples can be further processed according to the algorithms implemented with the Code Composer Studio environment. The similar scenario is utilized for transmitting processed samples from the DSP to the codecs incorporated in the extension card.

After D/A conversion signals are filtered employing low-pass, fourth-order, Butterworth filters in order to attenuate conversion artefacts. The extension card filters were designed with Texas Instruments Filter Pro application [33]. Relatively high order filters are required because codecs introduce significant distortion when the low sampling rate is chosen (e.g. 8 or 16 kHz) [34]. After the filtration analog signals are amplified using the Dclass power amplifiers and are presented to miniature loudspeakers through chinch connectors.

The DSP development board along with a dedicated extension card is a complete prototype of the contactless hearing aid. Furthermore, the emulation link between the Code Composer Studio environment and development board allows for efficient implementation of the algorithms because of the availability of all debug functions. Finally, algorithms incorporated in the contactless hearing aid may be evaluated in reallife conditions which is essential for their proper tuning. The final prototype of the contactless hearing aid can be easily designed in future by supplementing the extension card with the DSP processor itself and wit a booting memory.

Conclusions

In this paper a novel contactless hearing aid dedicated to infants is presented and its structure is thoroughly described. The reduction of the acoustic feedback is a major issue in this study, thus two different approaches are implemented. The first one utilizes the beamforming method based on an artificial neural network. The second one employs direct signal processing algorithm and forms additional protection against feedbacks. Two such techniques (adaptive notch filtering and adaptive feedback cancellation) were studied and the former one was chosen for experiments because it is more suitable for digital signal processors implementation.

In the experiments the nature of feedbacks has been examined and the algorithm for feedback component discrimination has been implemented. The results of the experiments prove that the algorithm is able to detect feedbacks with a good accuracy and it may be implemented on a digital signal processor.

The paper also presents algorithm implementation details and describes the dedicated extension card that together with a Texas Instruments DSP development kit will form a complex environment for the four channel audio processing.

Future work will be focused on evaluating the hardware prototype within the context of contactless hearing aid device effectiveness.

Acknowledgements

This work was supported by the Ministry of Science and Education within the Grant No. 3T11E02829.

References