Introduction
Epilepsy in itself is a brain disorder in which groups of nerve cells or neurons in the brain sometimes signal abnormally. In epilepsy, the normal pattern of neuron electrical activity becomes disturbed, causing strange sensations, emotions, and behavior, or sometimes convulsions, muscle spasms, and loss of consciousness. There are various things that can trigger an epileptic seizure such as flashing lights or sudden changes from dark to light (Photosensitive Epilepsy). Other people can react to loud noises or monotonous sounds, or even certain musical notes.
This project will focus on the development of a DSP-based system for a patient that suffers from recurrent bouts with sound induced epileptic seizures. Specifically, sounds that are produced by telephone (rotary, cellular) rings and pagers. It should be noted, however, that this patient has the ability to “block” the effect of the trigger by merely being aware that the ring is about to occur.
This project pursues the design of a real-time DSP system that could receive the sounds heard by the patient mentioned above with a microphone and deliver a processed audio signal to headphones worn by him, while;
1-) Providing an advanced warning of an incoming ring
2-) Attenuating the ring in the signal delivered to the headphones at least to a point where it will not trigger a seizure.
As such, the project pursues the design and implementation of the following systems:
I. Adapting Ring Detector:
This application of the system will track and detect any strong periodic component in an incoming audio signal and warn the patient of the impending ring with about 1 second ahead of time.
II. Ring suppression or cancellation:
The system will track, detect and attenuate these strong periodic components in the incoming audio signal. At this stage, attenuation rates of at least 25dB are being sought.
Subsequently, one of these two implementations will be used in the development of a stand alone real-time system that will utilize a dedicated DSP processor, such as the TMS320C31 or a TMS320C6201. The choice of which of the implementations will be used will ultimately be the decision of the patient.
We are currently favoring the implementation of the advance warning system, relying on the ability of the patient to “block” the negative effects of the ring, if he has advance notice of its impending occurrence.
Preliminary Solution
Our initial hypothesis modeled the ringing disturbance in the incoming signal as just a fixed periodic fundamental tone with its corresponding harmonics that was potentially superimposed on speech, or other audio components. According to this simple model of the targeted interference, an adaptive filter or ALP (Adaptive Line Predictor) could be applied to solve this problem. The ALP would be ale to separate the periodic and speech components from the incoming signal. The output of this filter would then consist of a relatively undistorted speech signal.
However, spectral analysis of a couple of types of this particular disturbance (cellular phone and office phone ringing sounds) brought about the notion that this disturbance is not just one strong fundamental tone, but rather a combination of more than one strong fundamental frequencies, all with their associated harmonics.
In addition, spectrogram of these rings shows that the spectral composition of these signals varies with time. Using a recorded waveform that contained speech and the ringing of an office telephone, an ALP was used to test the initial hypothesis. The ALP was able to remove some of the ringing, however, the power of this interference was still quite high, not to mention that severe degradation to the speech in the recording that also took place.
Adaptive Ring Detector:
This method will work around the shortcomings of the ALP in regard to its inability to deal with more than one strong, switching periodic component. In this method, the incoming signal spectrum will be broken down into several bands for subsequent processing by a set of ALP filters. Each filter will process its assigned band to determine if there is any strong frequency component present in that band.
By reducing the number of strong frequency components at the input of each adaptive filter this approach will enhance the performance of the ALP filters in terms of quickly detecting any strong periodic component that is present in the incoming speech signal.
Figure 2 shows a block diagram of the real-time adaptive system proposed. It includes several components, but it is simply a system that receives an input, delays it just enough to allow the hardware to perform the numerical calculations and then signal if at any time there is an impending ringing. As mentioned previously, since the focus is now mainly on detection, the system will give advanced warning of any incoming ringing by way of a visual indicator, such as a Light Emitting Diode (LED). As soon as any ringing or strong periodic component is detected in an incoming input audio signal, the LED would turn on and remain on for as long as that periodic interference is present.
Filter Bank:
The first block of the adaptive system consists of bank of eight IIR band pass filters, each having a bandwidth of 500Hz and covering a range from DC to approximately 10Khz. The upper limit of this range was decided upon based on knowledge of the hearing range for a normal human being. This is range is usually between 16 Hz to 16Khz, the upper limit falls off with increasing age. A bandwidth including up to 10KHz will satisfy speech intelligibility requirements. Each filter is a 6th order Chebyshev band pass IIR filter. The choice for this type of filter was made on the basis of the filter’s sharp frequency cutoff response. The output of each filter will isolate any significant strong spectral component that might be associated with the ringing or that could also be associated with any other type of input (i.e. sirens, beepers) or even speech.
Adaptive Line Enhancer and Detector:
Once the input signal has been partitioned into different spectral bands, each band is processed by its corresponding Adaptive Line Enhancer and Detector (ALED). The ALED can be thought of as consisting of two parts; an Adaptive Line Enhancer (ALE), and a modified spectral peak detector. The ALE filter will extract any strong periodic component(s) from its given spectral band. The idea here is that a periodic component is much more concentrated (larger power density) about a certain frequency than any portion of speech over the frequency spectrum in that assigned band.
The output of the filter is then compared with the input portion of the signal. The comparison is made on the basis of their RMS power measured over a small interval of time. Using the ratio of the filter output power over the error power (output_power / error_power), we can determine if there is a periodic component present by comparing it to a fixed threshold level. If the filter fails to detect any strong periodic component this ratio will be small, however in the event that the filter does detect some periodicity in the spectral band the ratio will be considerably larger. The boundary that separates these two events is set as a threshold value (Theta). If detection has taken place, a small pulse will be triggered in that assigned spectral band.
Ring Output Detector:
Given the fact that speech is quasi-periodic, there will be sections of speech that may contain a periodic component of considerable strength. However, we would like to prevent this speech periodicity from triggering a false alarm in the system. To address this problem, all the triggering pulses from the individual single-band ALEDs are fed into the Ring output detector, which will determine if there is ringing present. Since speech, at the most can only trigger a detect signal in two frequency bands, while a ring could trigger them in as many as four, all of the detect pulses are added up and the result is compared to a threshold value (Theta2). This value can be set to 2.1, and as such if the combined result of all the bands (detection pulse) add up to more than 2, then we can say that a ring has been detected.
Preliminary Results
The initial results of this project at this time have been obtained from computer simulations. It is hoped that by improving some of the critical areas of the system (i.e. speed of response, optimum Band Pass Filter length, ALE structure) in simulation, then the hardware implementation will not run into many obstacles.
Figure 6. Advance warning outputs are generated by the system, from a typical audio segment containing both speech and a telephone ringing.