@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix skos: . vivo:departmentOrSchool "Applied Science, Faculty of"@en, "Electrical and Computer Engineering, Department of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Chau, Kwok Wing Chau"@en ; dcterms:issued "2010-08-27T16:58:28Z"@en, "1989"@en ; vivo:relatedDegree "Master of Applied Science - MASc"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description """The objective of this project has been to design a reliable warning sound recognition system for hard of hearing and deaf people. Commercially available auditory warning devices use simple technologies, which are not able to produce the performance required. The demand for a versatile WARNing Signal Identification System (WARNSIS) that satisfies the needs of hard of hearing and deaf individuals has been well established. This WARNSIS must be "teachable" in order to cope with the many different sounds, and diverse noisy environments. Relevant sounds are telephone rings, sirens, and smoke and fire alarms, and noise includes all other sounds including radio-music, conversation, machinery, etc. In the absence of published data, we studied extensively both timing and spectral characteristics of warning sounds. We found that the average short-time absolute amplitude of warning sounds is useful in providing timing information, and that the short-time spectra yield characteristic patterns for signal classification. The WARNSIS operates in real-time, and embodies two parts: the timing analyzer and the spectral recognizer. The timing analyzer continuously monitors the variations of environmental sounds, from which important timing features are derived. If a potential warning sound is detected, the spectral recognizer is activated to analyze its spectral patterns. When these patterns match one of the learned and pre-stored templates, a warning sound is identified with the known warning sound associated with that template. An advantage of such a recognition scheme is that it avoids unnecessary and computationally intensive spectral analysis work when only noise is present. Evaluation results show that the WARNSIS can reliably recognize warning sounds in random noise with no false alarms. In loud music and conversation backgrounds the WARNSIS can still achieve a high recognition rate, but more false alarms are generated. In household environments where conditions are less demanding than our evaluation criteria, our system is expected to produce very satisfactory results. Since the WARNSIS can be taught to learn and recognize new warning sounds, it may be used in other applications such as noisy industrial sites and traffic light control."""@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/27831?expand=metadata"@en ; skos:note "A W A R N I N G S I G N A L I D E N T I F I C A T I O N S Y S T E M ( W A R N S I S ) F O R T H E H A R D O F H E A R I N G A N D T H E D E A F K w o k W i n g Chau B . A . Sc. Universi ty of Windsor A THESIS SUBMITTED IN PARTIAL F U L F I L L M E N T OF T H E REQUIREMENTS FOR T H E D E G R E E O F M A S T E R OF A P P L I E D SCIENCE i n T H E F A C U L T Y OF G R A D U A T E STUDIES D E P A R T M E N T OF E L E C T R I C A L E N G I N E E R I N G We accept this thesis as conforming to the required standard T H E UNIVERSITY OF BRITISH COLUMBIA July 1989 © K w o k W i n g C h a u } ] 989 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of EC CCr/ttCti*- £AJ Cfo £~&//J 73 dB, relative to 1 pW. The second major component shall have a mean sound power level of > 68 dB, relative to 1 pW; 2. The total mean acoustic power level shall be > 80 dBA, relative to 1 pW. These power levels apply with the volume control set for maximum volume; 3. At least one of the major component (fl) shall be below 2000 Hz. The nominal frequency of the higher major frequency component (f2) shall be equal to or greater than 5/4 of the lower major frequency component (fl), i.e., f2 > 5/4 fl; 4. The alerting signal of a telephone with an electronic alerting device that does not produce an acoustic spectrum rich in overtones shall meet the criteria in 1), with the exception that f l and f2 shall each have a mean power level of > 73 dB, relative to 1 pW; Chapter 2. Warning Sounds and Generating Devices 12 5. A telephone shall have a loudness adjustment accessible to the user that produces at least of a 6 d B A total attenuation when operated from its high to low volume position; and 6. W i t h regard to r inging cycles, r inging current supplied by telephone company central office shall belong to one of the following sequences : • Repeti t ive bursts of 2 seconds out of every 6 seconds where an individual burst may be as short as 0.8 second; • Repet i t ive bursts of 1 second out of every 4 seconds where an individual burst may be as short as 0.6 second; or • Repeti t ive bursts of at least one r inging burst of a m i n i m u m 0.5 second durat ion i n any 4 second period. 2.3.2 Smoke Detector A l a r m Sounds Smoke alarms are used to alert people to the presence of smoke and to the potential of fire. Generally, this warning sound is very strident and insistent. In a study of a la rm sound attenuation inside residential buildings Hal l iwel l and Sul tan [13] investigated the spectral content of the sounds produced by a number of smoke detectors. Using a 2-channel F F T analyzer connected to two microphones, they obtained the short-t ime spectra, and for each sound 64 of these short-time spectra were averaged to give the spectrum. T h e narrow-band spectrum was subsequently converted to a th i rd-octave spectrum by simply summing the energy wi th in the third-octave bands. Their results for various smoke detectors show two or more strong spectral components in al l computed spectra [Table 2.1]. Unfortunately, this work did not include the investigation of the var iat ion of the short-time spectra obtained from consecutive samples. Chapter 2. Warning Sounds and Generating Devices Table 2.1: Spectral analysis results for different smoke detectors [13] Detector Typet 1/3 Octave Frequency Bands (kHz) 0.5 0.63 0.8 1.0 1.25 1.6 2.0 2.5 3.15 4.0 5.0 A l 38* 39 39 39 63 57 73 96 84 63 50 A2 37 38 38 38 44 56 70 98 92 67 56 BI 82 82 60 71 74 81 79 95 95 95 88 B2 79 81 66 72 76 81 77 93 94 96 92 CI 44 44 44 45 45 50 61 79 102 90 69 C2 44 44 44 45 45 50 62 79 102 91 70 D l 46 46 46 46 47 52 63 80 103 93 71 D2 44 44 44 45 45 50 62 80 102 88 68 E l 84 70 69 85 76 92 88 96 92 91 80 E2 76 83 63 69 80 87 85 97 100 91 89 F l 61 60 72 70 70 74 86 75 83 90 82 F2 58 61 69 70 72 77 90 81 82 89 82 G l 37 37 37 . 38 39 50 63 88 95 69 55 G2 38 38 38 38 39 48 61 84 95 71 56 t : Detectors with same letter denote identical model. $: Maximum Sound Power Output in dB Chapter 2. Warning Sounds and Generating Devices 14 2.3.3 W a r n i n g and A l a r m Sounds Generated by Vehicles and Traffic C o n -trol Devices Miyazaki and Ishida [14] have studied the spectral characteristics of traffic alarm sounds commonly used in Japan. Such include sounds produced by electric horns used in passenger cars, small, middle size buses and trucks; air horns used in large buses, heavy duty trucks, and trailers; sirens used in emergency vehicles; horns used in rail-road crossing; and traffic noises. Their observations have only limited value for us since they neither give description of the techniques used nor do they specify the type (short-time or long-time aver-age) of the spectra obtained. Table 2.2 summarizes their results. They conclude that traffic-alarm-sounds have sharp line spectra, whereas ambient traffic noise is wide-band random noise. Table 2.2: Summary of spectral analysis results for traffic alarm sounds [14] Traffic Alarm Devices Installed Vehicles Major Frequency Features Electric horn Passenger cars, small, middle size busses trucks basic resonant frequency at 300 Hz - 500 Hz, dominant harmonics at 2.0 - 4.0 kHz Air horn large busses, heavy duty trucks, trailers dominant peaks at 300 - 500 Hz Siren Emergency vehicles dominant peaks at 700 - 2000 Hz Rail-road crossing 2 - 3 dominant peaks at 2.0 - 4.0 kHz ambient traffic noise broadband noise below 300 Hz In British Columbia, and typically in North America, three types of emergency ve-hicle siren sounds are used: the \"hi/lo\" sound, the \"yelp\" sound, and the \"wail\" sound. Chapter 2. Warning Sounds and Generating Devices 15 The h i / l o sound is usually found on most ambulances. It consists of two alternating tones, and w i t h the pattern repeating about once per second. T w o commonly used tone pairs are 690/920 H z and 520/1520 H z . The wai l sound is a slow changing tone between two preset tone frequencies. A typical example is the wai l sound used by police motorcycle sirens w i t h preset tone frequencies at 500 H z and 1460 H z , and a repetit ion rate of 10 cycles per minute [15]. The yelp sound is a fast changing tone between two preset tone frequencies. A typical example is the electronic siren produced by Southern Vehicle Products Inc., which provides a yelp sound wi th preset tone frequencies at 600 H z and 1350 H z , and a repetit ion rate of 3 to 5 cycles per second [16]. The yelp and wai l sounds are used by both fire-trucks and police cars. 2.4 T h e Emerg ing Scientific basis for Generating Warn ing Sounds W h i l e warning sounds have been used for a long time, many of these are based on subjective opinions as to what is \"best\". On ly recently was any scientific work done to determine what sound characteristics w i l l elicit opt imal responses under varying circumstances. Such work is part icularly relevant for us, since i n the future warning devices may follow a more systematic approach to sound generation than it has been the case un t i l now. 2.4.1 A Generic Warn ing Sound Generating Scheme Accord ing to the work of Patterson and his colleagues, a warning sound need not to be excessively loud, but its amplitude must depend on the background noise level. They have demonstrated, that i n order to hear sounds reliably in noise, some spectral components must be between 15 d B and 25 d B above the masked threshold [17,18]. Lower and Wheeler [19] has developed a desk-top computer program to estimate this Chapter 2. Warning Sounds and Generating Devices 16 background threshold. W i t h the estimated background threshold, the spectral compo-nent amplitude of the warning sound can be determined. Th i s approach had been used to study the intense background noise of mi l i ta ry helicopters i n the U . K . [20]. W i t h regard to the frequency content of the warning sound, Pat terson [17] l imits it to the range between 0.5 k H z and 5.0 k H z . Based on these spectral amplitude and frequency l imits of the warning sounds, a pattern of pulsative sounds which is distinctive and resistant to undesirable noise contamination, was constructed by Patterson [17,22,23]. A s shown i n F i g . 2.1 , this prototype warning sound basically consists of a sequence of bursts each of which is made up of a sequence of pulses. Different degrees of perceived urgency can be manipulated by simply varying the characteristics of the pulse sequences. In Patterson's work, the pulse design starts w i t h measurement of the ambient noise spectrum. Then , the warning signal spectrum is determined by setting al l its compo-nents 15 - 25 d B above the corresponding ambient noise spectral values. In order to avoid excessive peak factors i n the signal waveform, sine or cosine phase is assigned to the spectrum. Consequently, the pulses are generated by applying the Inverse Fast Fourier Transform. These pulses vary i n durat ion from 75 msec to 200 msec i n accor-dance w i t h the guidelines set down by Patterson [17,23]. A l so , the pulses are gated w i t h sinusoidal ramps at both ends i n order to avoid uncontrollable transients. A t this stage, by varying the fundamental frequency, and the relative weight of high and low frequencies of the pulses, any degree of perceived urgency can be designed. Usu-ally, greater urgency is signalled by higher fundamentals, and by relatively more high frequency energy. A burst is produced by assembling three-to-nine copies of the basic pulse. B y changing the elapsed time between the start of one pulse, and the start of the next, distinct p i tch and temporal patterns may be created. B y varying the amplitude of Chapter 2. Warning Sounds and Generating Devices 17 the pulses different loudness patterns may be obtained. The perceived urgency is generated by changing the overall p i tch, the speed and the loudness pattern of the pulses. In general, a burst w i th a high pulse rate w i l l convey greater urgency than a burst w i t h a low pulse rate. A rising pitch-contour can produce a more urgent burst than a falling pitch-contour. Addi t ional ly , an urgent burst w i l l remain at, or near, the m a x i m u m loudness while a less urgent burst w i l l decrease in loudness towards the end of the burst. Such bursts serve as templates from which warning sounds may be synthesized. The amplitude variations and spacing of the bursts are determined experimentally. The cri terion is that the resulting warning sound should effectively convey the desired specific warning message to personnel i n the v ic in i ty without act ivating their startl ing reflex. Pat terson successfully implemented this scheme on warning systems of commercial aircrafts and mi l i ta ry helicopters [17]. A slight modification of this scheme was also adopted for medical equipment used in intensive-care units and operating theatres of hospitals i n the U . K . [22,23]. Time in seconds AUDITORY WARNING SOUND COMPONENTS Figure 2.1: Auditory Warning Sound Components [17,21,23]. Chapter 3 Measurement and Analysis of T i m i n g &c Spectral Characteristics A s we have seen i t i n Chapter 2, the literature on warning sounds yields l i t t le useful information on their t iming and short-time spectral characteristics. Since i t is the purpose of this work to apply t iming and short-time spectral analysis techniques to systematically extract the unique identifying characteristics of these warning sounds i n real-life environments, such information is essential for us. Specifically, the detailed knowledge of warning sound characteristics provides the basis for the exploration of different signal recognition schemes. 3.1 T i m i n g Characteristics The objective of this part of our work was to derive useful information on the t iming of warning sounds from measurements of signal waveforms. For this purpose we used telephone rings, siren sounds, and smoke a larm sounds. Telephone rings were generated by bo th electro-mechanical and electronic ringers; siren sounds were produced by an electronic siren driver; and the smoke a la rm sounds were obtained from a commercial smoke a larm. 3.1.1 A P C - B a s e d D a t a Acquis i t ion System To obtain quantitative data, a PC-based data acquisition system was designed and constructed. Th i s system accepts the instantaneous absolute amplitude waveform of the signal, and transforms it into the short-time average absolute ampli tude ( S T A A A ) 19 Chapter 3. Measurement and Analysis of Timing &: Spectral Characteristics 20 waveforms. T h e n , the transformed waveforms are stored for plot t ing. The instanta-neous ampli tude and the short-time average variations i n absolute amplitudes of the signal are given i n Table 3.3, where x(n) represents the discrete instantaneous signal amplitudes, and N denotes the number of samples accumulated. Table 3.3: Instantaneous and short-time signal amplitudes signal amplitudes absolute signal amplitudes instantaneous short-time average x(n) 1 N J V n=l \\x(n)\\ J V n=l The instantaneous absolute signal amplitudes are generated by hardware, and the derivation of the short-time average absolute signal amplitudes, and storage of these derived samples is accomplished by software. F i g . 3.2 shows the block diagram of the method used to generate the discrete instantaneous absolute signal amplitudes. Basical ly, sounds are collected by a suitable microphone, are pre-amplified by a low-noise voltage amplifier, and are low-pass filtered prior to input to a full-wave rectifier. The output from the full-wave rectifier gives the instantaneous amplitude of the waveform. Then , an 8-bit A / D converter samples this waveform at 10 k H z . Consequently, the digitized sample is stored temporari ly in an output buffer un t i l the 8-bit microprocessor ( I N T E L 8088) is ready to accept the data v i a a bi-directional bus. In addit ion, a L E D bar graph is used to display the variations i n the instantaneous absolute amplitudes of the signal waveforms. Chapter 3. Measurement and Analysis of Timing & Spectral Characteristics 21 MICROPHONE SIGNAL PRE -PROCESSOR FULL-WAVE RECTIFIER CPU A/D CONVERTER Figure 3.2: Signal acquisition and derivation of instantaneous absolute signal ampl i -tudes In this implementat ion, the short-time average absolute signal amplitudes are de-rived from 12.8 msec accumulation of the instantaneous absolute signal ampli tude sam-ples ( A / D converted data). W i t h these instantaneous signals sampled at 10 k H z , a sample of the short-time average absolute signal amplitudes can be obtained by sum-ming 128 of the instantaneous signal samples. In order to avoid the problem of overflow dur ing the accumulat ion process, a 16-bit register is used to accumulate this sum. C o n -sequently, a sample of the short-time average absolute signal amplitudes is obtained by d iv id ing the 16-bit register content by the total number of accumulated samples (i.e 128 i n this case). The resulting quotient is then rounded to eight bits to provide the short-time average absolute signal amplitude sample which is transferred to a desig-nated file. T h i s file stores 1000 bytes. These data manipulat ion and transfer procedures are repeated unt i l the data file is completely filled w i th 1000 samples (equivalent to Chapter 3. Measurement and Analysis of Timing ic Spectral Characteristics 22 12.8 sec of the signal waveform). The program to handle this data manipulat ion and transfer i n real-time was wri t ten in I N T E L 8088/8086 assembly language. A flowchart of these operations is shown i n F i g . 3.3. 3.1.2 D a t a Collection W i t h this data acquisit ion system, we collected data on the absolute amplitudes of warning sounds i n the normal acoustic environment of our laboratory. F i g . 3.4 shows the experimental set-up. The siren horn produced siren sounds; and a radio cassette player provided the pre-recorded telephone rings and smoke a la rm sounds. A sound pressure level (SPL) meter placed aside the microphone measured the S P L variations of the environment throughout the data collection process. The S P L meter was set to \" C \" weighting and \" S L O W \" response, because the \" C \" weighting network of the S P L meter has a flat frequency response similar to that of the signal processing circuit of the data acquisit ion system; and the \" S L O W \" response provides an average of 1.0 sec of the acoustic energy variations of the environment. Based on the S P L measurements i n the absence and during the presence of warning sounds, the signal-to-noise ratio (SNR) could be deduced. S N R , i n this work, is defined as the ratio of peak signal power to peak noise power. Noises, i n this thesis, are defined as al l sounds other than warning sounds. Such unwanted sounds may include steady and transient random noises, radio broadcasts, or surrounding conversations. A detailed derivation of the relationship between the S N R and S P L measurements is given in A p p e n d i x A . D a t a on absolute amplitudes of warning sounds were collected i n two different back-ground environments. The first set of data were collected i n a steady random noise background which originated from a venti lat ion fan of a PC-computer . Such noise is typica l for office environments. A value of 60-62 d B C was recorded throughout the data Chapter 3. Measurement and Analysis of Timing Sc Spectral Characteristics t S T ^ B T J 1=1. j=1 • I INPUT E N D _ O F _ C O N V E R S I O N (EOC) S T A T U S F R O M A/D A C C < ~ - A C C + D , A C C < — - ACC/16 Xj < -- A C C 1 • J < — J + 1 | Y E S S T O R E X T O A S P E C I F I E D FILE Figure 3.3: Flowchart of procedure to accumulate and store 1000 sampl Chapter 3. Measurement and Analysis of Timing < f c Spectral Characteristics CASSETTE PLAYER AND RADIO Figure 3.4: Experimental set-up for data collection Chapter 3. Measurement and Analysis of Timing &: Spectral Characteristics 25 collection process. This set of data were referred to as \"clean\", because the SNR was maintained at least over 20 dB. To study the effect of more complex background noises on warning sounds, the second set of data were collected at a SNR of 10 dB. The back-ground noise sources consisted of both steady random noise, radio music broadcast, and speech. To establish the short-time average absolute amplitude profiles of the various noise sounds (without warning sounds present), a third set of data was also collected. This included all the noise sources used above, and the noise SPL was the same as that used in the SNR measurements. 3.1.3 T i m i n g Features of Different W a r n i n g Sounds The plots of the first set of data are shown in Fig. 3.5, Fig. 3.6, Fig. 3.7 and Fig. 3.8. Since the purpose of these measurements is to establish the time variations of the short-time average absolute signal amplitudes, the actual value of these amplitudes is of no particular interest. Therefore, the vertical axes show a relative scale without units. The following observations may be drawn from these figures: 1. Fig. 3.5, Fig. 3.6(b) & (d) (siren sounds), and Fig. 3.7(a) &: (b) (telephone rings) show on-off type repetitive patterns of warning signal bursts; Fig. 3.6(a) &: (c) (siren sounds), and Fig. 3.7(c) (smoke alarm sound) display the steady sounds; 2. Fig. 3.5 (a) and (b) show devices which produces sounds with very similar tem-poral structures, but with different repetition rates; 3. Fig. 3.5 (d) is a two-tone siren sound, and its amplitude contour can be charac-terized by i) a transition from background level amplitudes, and ii) a repetitive Chapter 3. Measurement and Analysis of Timing &: Spectral Characteristics 26 on-off pattern representing two tones of different intensities (for other siren sounds or telephone rings, the off-patterns represent the background noise levels); 4. The width of the bursts of these waveforms varies from 102.4 msec to 3.24 sec; 5. The repetition period of on-off patterns ranges from 140 msec to 5.86 sec; 6. Steady sounds are characterized by signal level transition to higher steady am-plitude level; and 7. Contours of the average of short-time absolute signal amplitude of radio broad-casts (Fig. 3.8) consist of random, nonrepetitive sequences of signal bursts. The plots of the second set of data are shown in Fig. 3.9 and Fig. 3.10. Comparative examination of these plots yields the following observations: 1. For short-burst, such as (a), and (b) in Fig. 3.9, and (d) in Fig. 3.10, the introduction of radio broadcast background alters the baseline levels, and smooths out the weak peaks of the \"clean\" signals; however it produces no significant change in relative timing between consecutive amplitude peaks of the waveforms; 2. For signals with long silence intervals( > 400 msec) such as (c), and (d) in Fig. 3.9, and (b) in Fig. 3.10, spurious small peaks appear randomly during these intervals; and 3. The repetition rate of the on-off patterns of burst-type sounds is unchanged by variations in background noise. In summary, we can conclude from these measurements that the short-time average absolute amplitude contours provide unique timing information on both steady and burst-type sounds. Chapter 3. Measurement and Analysis of Timing & Spectral Characteristics i i i i 1 1 1 i i i 1 1 1 i 0.000 0512 1 024 1 538 2.048 2 580 0.00 2.38 5 12 7.88 10.24 12.80 TIME ( in sec) TIME ( in sec) (b) (d) Figure 3.5: Short-time average absolute amplitudes (STAAA) of siren sounds: a) Jl : Burglar alarm (JDS-100); b) J2 : MPI-11; c) J3 : JDS-100 I; and d) J4 : HI-LO Chapter 3. Measurement and Analysis of Timing & Spectral Characteristics 28 0.0 I™ rTTTTinnrnrr — i — i.a ~ i — 3.2 TIME ( i n sec) (a) \"I 4.8 8.4 0.0 Jfirrmrnr —r— 1.8 — I — 3.2 TIME ( i n sec) (c) - i — 4.8 - I 0.4 0.000 0.912 TIME ( in sec) (b) 0 000 0 256 0.512 0.768 1 024 TIME ( in sec) (d) 1.280 Figure 3.6: Short-time average absolute amplitudes (STAAA) of siren sounds: a) J5 : High steady sound; b) J6 : Pulser; c) J7 : Steady horn; and d) J8 : Electronic Synthesized Bell sound Chapter 3. Measurement and Analysis of Timing