Sleep Inc. ANC / Anti-Snoring Technology Overview

This overview was prepared by Obsidian technology for  Sleep Inc., a customer of this product.

The theory behind Active Noise Cancellation Technology (ANC) is well established and in resent years consumer products have appeared which utilize basic analog technology for reduction of machinery and aircraft noise. An example is the Bose ANC headset.


More complex, digitally based, ANC devices have been mostly limited to industrial and commercial applications. However, the continuing reduction in cost of Digital Signal Processing (DSP) components now makes it possible to introduce more complex digital ANC products into consumer markets.
 

Digital ANC systems use adaptive digital filters to make anti-sound as shown in figure 1. When sound and anti-sound are combined at the listener's ear, the sound is cancelled. The electronics take advantage of the fact that electronic signals travel much more quickly than sound. Therefore noise signals picked up by a remote microphone can be processed into anti-sound before the noise arrives at the listener. 
 

Practical ANC systems generally require at least two channels: one for each ear. Providing a microphone close to each ear is important for good performance since the system needs to know what the listener hears in order to train. Traditionally this has been difficult to do wirelessly because the technology was too heavy and too expensive. However, high volume wireless headset technologies are being developed (E.g. for cell phones) which are small, light, and have excellent battery life. Low cost processing technology and low cost wireless technology make high volume ANC systems practical.


Basic Digital ANC Technology


Digital ANC systems function by learning (or training to) the acoustic environment. The learned acoustic environment may then be employed to produce the anti-sound. Microphones near the listener's ears provide the error information necessary for characterizing both the room environment and the imperfections in the system components. This is a complex process since rooms generally have very complex sound reflection and absorption characteristics.

However, the basic room acoustic environment can be learned by using a Least Mean Square (LMS) digital adaptive filter as shown in figure 2.
 

Figure 3 shows the result of capturing a response in this way. This represents this "impulse response" of the room over 1/8th of a second: I.e. it represents the variation in sound pressure in the test room after an impulsive noise event (like a gun shot). 

 
Figure 4 (left) shows the main components of a typical single channel ANC system. The basic FIR filter uses filter coefficients 'H' that represent the opposite of the room response.

Signals are converted between analog and digital domains by Analog to Digital Converters (ADC), and back to the analog domain by Digital to Analog converters (DAC). 

Most of the complexity of the system is in the update system that maintains the value of H at system startup or if there is a change in position of the noise or the listener.
 

System Training


Digital ANC systems must perform two different types of system training:
1. The basic room acoustic response H.
2. The response of the ANC plant (X) that defines the performance of the correction system. I.e. the path from mic1 to spkr1. 
The room acoustic response H is trained by observing the response of mic and mic1 to the noise (in this case the snore). However, for this to work properly the plant, or X response, has to be trained first.

X plant training schemes are of two basic types1: Static Training in which X is trained at system startup and Dynamic Training in which X can be continuously trained during the operation of the system. Dynamic training has the advantage of allowing subjects to move after system start.

In either case some artificial training sound is needed to characterize the plant X. Plant training is generally achieved using a system like figure 2. Fortunately, the correlation features of the LMS filter mean that uncorrelated sounds (i.e. sounds not related to the artificial noise or training sound) are rejected in the resulting X response. Hence, the level of artificial sound required from the speakers to keep X trained can be very low such that for the listener it is close to the threshold of hearing.



Limits to ANC Performance


The physical characteristics of the audio environment place some fundamental limits on the performance of typical ANC systems irrespective of the quality of the electronic components. These are:

 
1. Causal Distance. As shown in figure 5 the ANC microphones must be placed such that there is time to process the signal into anti-sound. This cannot happen if the noise source and the listener are too close. Note that the distance C can be optimized if an array of low cost microphones is used. In this case the DSP processor chooses the microphone that is closest to the snorer. Another option is to place the snorer microphone directly on the snorer.

 
Figure 5. Causal Distance
2. Spatial Resolution. For perfect cancellation the noise and anti-sound wave must exactly cancel at a point in space. In practice ears, microphones, and speakers are distributed in space. This makes it impossible to achieve perfect cancellation over a wide range of frequencies. In addition the microphones and speakers are always moving slightly and are not at the ideal position. This also limits the maximum achievable attenuation and attenuation bandwidth. 
3. Error Microphone to Ear Distance. In a practical configuration the error microphones cannot be placed on the eardrum. The distance between them means that perfect cancellation is not possible because the system cannot train to exactly what the eardrum hears. 


This effect was measured in figure 6 with a model ear and an ANC system working with white test noise. There is a progressive loss of ANC performance with frequency which reaches 8dB at 3KHz. For this reason few ANC systems attempt attenuation above 2KHz.

Snoring Characteristics


Much work has been done on the characterization of snoring sounds 2,3. Fortunately most sound energy is below 2KHz where ANC is most effective.

Figure 7 (right) shows the frequency spectrum during a typical snore event. However, there are significant variations in spectrums between snorers. The other major characteristic of snoring is it's episodic nature. Snores are separated by silences which vary from a fraction of a second to hours. The non-continuous nature of snoring can be used to simplify the system. Note that while the snore spectrum of figure 7 is typical, some snores have higher frequency content that cannot be reduced by conventional ANC. These are sounds primarily produced by the whistling of air through the teeth and lips.
Figure 7. Spectrum of Snoring.
 

Practical Systems

A more practical ANC system for snoring is shown below in figure 8.

This uses two channels, one for each ear. The microphones and speakers are built into a comfortable headset. This architecture is typical for personal quietness systems. 

Figure 8. System Diagram
At power up the X plant parameters must be trained. This is achieved with the system shown in figure 9 that runs at system reset before the ANC commences. It is, in effect, a two-channel version of figure 2. X parameters X1, and X2, kept updated by this.

Figure 9. Plant training.
A measurement of the development system yielded the results of figure 10 (right). This achieves around 15dB noise attenuation over the 200Hz to 2000Hz frequency range. 

Development hardware combines a number of standard DSP, digital and analog components. The TI TMS320C5410 processor provides DSP processing.
Figure 10. Performance.

 

Masking Algorithms

Basic ANC does not remove the higher frequency (above 2KHz) produced by snorers. To cover (or mask) these residual noises masking sounds can be produced by the system at the users option.

Products that produce masking sounds have an established market. An example is the Sharper Image Travel Smoother® 20. This provides large selection of soothing noises such as rain and breaking waves designed to cover background noises and provide an acceptable sleeping environment.

However, if a masking sound is acceptable to the user, it is possible to actively modify the characteristics of the masking sound such that residual sounds are further reduced in perception. Algorithms that perform this task take advantage of the perceptual characteristics of the human mind. Namely:

1. Mammals have evolved to focus on discontinuous sounds since these are associated with danger, food, or social communication.
2. Changes in pitch during sound events trigger innate responses to vocalization in the human brain that are undesirable while attempting to sleep.

Sleep Inc. is exploring proprietary algorithms which both attempt to keep the total amplitude of the received noise constant over time and to continuously whiten the sound spectrally. Patent applications are expected in this area.



Patent

Sleep Inc. has title to US patent #3,998,209 that covers the application of ANC to the problem of snoring. This patent describes the system shown in figures 8 and 9 and details the physical arrangement of microphones and speakers in the vicinity of a domestic bed.



Heuristics of the Design

Low level Training Sounds. The sound used to train the plant cannot be disturbing to the user. Here we can take advantage of the correlation properties of the LMS. A very low- level pseudo random sound is played through the users speakers. Sounds picked up by the microphones that are correlated with the pseudo random sound cause training. Other sounds are rejected even if they are at a higher level.

Figure 11. Snore Detection.		 Snore Detection. 

To prevent false training of the main LMS during quite (non snoring) periods an algorithm is needed to accurately determine when there is sound coming from the snorer. The heuristically developed algorithm for this is shown in figure 11.

 
 

Normalized LMS. A basic LMS algorithm is sensitive to the level of the snore. In particular the training time of the snore increases as the snore volume decreases. Hence the LMS µ is modified on the fly to be inversely proportional to the average snore level.


Code Portability. The DSP software is organized for best compromise between portability and performance. The top-level code and much of the control code is written in C. Assembly code is only used for simple DSP function blocks such as the LMS and FIR filters. These blocks are also available for most DSP architectures as library elements. The CPU spends more than 90% of its time in LMS and FIR elements. Hence the loss of efficiency is minimal with this mixed more approach.
 

Practical System Arrangement: 

Figure 12		 Figure 12 shows the general arrangement of speakers and microphones in a bedroom setting.

References


1. S. M. Kuo and D. R. Morgan, Active Noise Control Systems: Algorithm and DSP Implementations. New York: Wiley, 1996.
2. Sleep Solutions US Patent 3,998,209. 
3. F. Dalmasso at el, Digital Processing of Snoring Sounds. 
4. A. Karakasoglu, C. Hung, J.F. Abbott, and S.C. Douglas, A Low-Cost Multichannel Active Noise Control System for Personal Quietude, Proc. 29th Asilomar Conf. on Signals, Systems, and Computers, Pacific Grove, CA, vol. 2, pp. 1275-1279, November 1995.