Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
Augmented warning sound detection for hearing protectors Eric R. Bernstein, Anthony J. Brammer, and Gongqiang Yu Ergonomic Technology Center, University of Connecticut Health Center, 263 Farmington Avenue, Farmington, Connecticut 06030-2017 [email protected], [email protected], [email protected]
Abstract: Perception of warning sounds, such as vehicle backup alarms, is reduced when hearing protection devices (HPDs) are worn. A crosscorrelation approach is employed to detect a pre-selected warning sound and enable it to bypass the attenuation of the HPD while still attenuating the environmental noise. Computer simulation shows that the algorithm can detect the specified alarm at signal-to-environmental-noise ratios as low as 30 dB. Human subject testing of the algorithm, implemented on a modified commercial HPD, confirms the minimum detection threshold obtained in simulation, and demonstrates a 7 dB improvement in detection threshold compared with the unmodified HPD. C 2013 Acoustical Society of America V
PACS numbers: 43.66.Vt, 43.60.Bf, 43.60.Mn [QJF] Date Received: October 1, 2013 Date Accepted: November 14, 2013
1. Introduction In high sound pressure level (SPL) environments, the perception of alarm signals is reduced when hearing protection devices (HPDs) are worn (Edworthy and Hellier, 2000). Two considerations are of importance in these situations: alarm detection and alarm localization (Casali et al., 2004; Alali and Casali, 2011). This paper explores a method for augmenting alarm signal detection when using HPDs in noisy environments. The proposed algorithm searches for a pre-selected alarm signal in the environment and, if found, allows that alarm signal to bypass the passive attenuation of the HPD while maintaining attenuation of the environmental noise. The normalized correlation-based approach used in this detection algorithm is similar to methods proposed by Lutfi and Heo (2012) and Carbonneau et al. (2013). However, these methods rely on the periodicity of the autocorrelation function for detecting a general alarm signal, rather than optimizing the detection of a specific alarm using the cross-correlation with the expected signal. The performance of the proposed algorithm has been evaluated by computer simulation as well as when applied to a modified commercial HPD worn by human subjects in environmental noise. 2. Methods Figure 1 shows a block diagram of the proposed alarm signal detection algorithm integrated into a circumaural hearing protector. The system uses an external microphone, E, mounted on the surface of the ear cup to detect the unwanted environmental noise mixed with the alarm sounds. The microphone signal, e(n), is passed through a filter, Hebp(z), to remove unnecessary frequency components that are outside the expected range of the tonal warning signal. For such warning signals, Hebp(z) would take the form of a bandpass filter with the passband centered around the expected alarm signal frequency. The absolute value of the resulting bandpassed alarm signal abp(n) is passed through a lowpass envelop filter, Hen(z), to generate aen(n), which is then downsampled by a factor of J. For this study, the sampling rate from the external microphone was set to 24 kHz and J was set to 96, resulting in a 250 Hz sampling rate for the alarm detection algorithm. The N most recent output samples of the downsampler are then
J. Acoust. Soc. Am. 135 (1), January 2014
C 2013 Acoustical Society of America V
EL29
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
Fig. 1. Block diagram of proposed alarm signal detection algorithm implemented in a circumaural hearing protector.
stored in the vector r(j) for use in the correlation stage of the algorithm. Before using r(j) for alarm detection, the vector must be normalized to account for the wide dynamic range of sounds in the real world. The normalized vector rn(j) is given by rn ðjÞ ¼
rðjÞ r ðjÞ ; rr ðjÞ
(1)
where r ðjÞ and rr ðjÞ are the mean and standard deviation of r(j), respectively. The normalized signal vector is then cross-correlated with a length N basis vector, s, to generate the detection signal, d(j), according to N 1 X dðjÞ ¼ sðiÞrn ðj iÞ:
(2)
i¼0
The coefficients of the vector s are tuned to the characteristics of the alarm signal that the algorithm is designed to detect. The s vector coefficients are determined off-line and are given by the vector rn(j) from Eq. (1) under the condition when the alarm signal is present with no environmental noise. The detection signal, d(j), is then used in the “Threshold” block to generate a binary signal, T(j), that takes a zero value if the alarm is not present and a unity value if the alarm signal is present. T(j) is thus found according to 1 if dðjÞ D or m > 0; TðjÞ ¼ (3) 0 else; where D is the threshold parameter and m is a counter decremented at each algorithm iteration to ensure T(j) maintains a value of unity for some time after a threshold has been exceeded. For this purpose, m is set to the value of a count parameter, M, when an “alarm” has been detected. This enables a suitable gating signal to be generated for gain control. Once a signal has been detected, the system must alert the user to the presence of the alarm. In this paper, the alarm signal is presented to the user at the SPL that occurs in the environment, and thus the alarm signal (but not the environmental noise) should bypass the attenuation of the passive ear cup. To accomplish this, the bandpass
EL30 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
filtered signal abp(n) is passed through a variable gain block G(j) to generate the output signal y(n) for loudspeaker, L, located underneath the ear cup. In this way, the alarm noise from outside the ear cup is reproduced underneath the ear cup, while the environmental noise outside the narrow passband of abp(n) remains attenuated. To establish the appropriate gain, it is first necessary to determine the environmental noise SPL underneath the ear cup. For this purpose, the sound generated by the loudspeaker must be removed from that sensed by the internal microphone, I[i(n)]. An estimate of the algorithm-generated warning sound to subtract from i(n) is obtained by passing y(n) through a transfer function model from L to I, which represents the electro-acoustic transfer function of the loudspeaker and microphone as well as the sound propagation from L to I. The resulting signal y0 (n) is subtracted from i(n) to yield i0 (n). The signal i0 (n) is then passed through a bandpass filter, Hibp(z), to generate ibp(n), which is used to measure the SPL of the noise underneath the ear cup. The SPL estimates of abp(n) and ibp(n), namely, A(n) and I(n), respectively, are found according to AðnÞ ¼ b Aðn 1Þ þ ð1 bÞjabp ðnÞj2
(4)
I ðnÞ ¼ aI ðn 1Þ þ ð1 aÞjibp ðnÞj2 ;
(5)
and
where a and b are exponential forgetting parameters usually taking values close to unity. If an alarm signal is not detected, G(j) takes a value of zero, thereby disabling electronic transmission of sound from the environment to within the ear cup. If an alarm signal is detected, the gain should be such that the alarm signal underneath the ear cup has the same SPL as outside, and thus the gain must take into account the attenuation of the earmuff. The gain G(j) is then 8 > 0 if TðjÞ ¼ 0; > < sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (6) GðjÞ ¼ AðjÞ I ðjÞ > if TðjÞ ¼ 1: > : AðjÞ þ c Using the gain selected, the signal underneath the ear cup is set so that the summation of the alarm sound generated by the loudspeaker and alarm sound as attenuated by the HPD should maintain the same SPL as the alarm signal outside the ear cup. To evaluate the effectiveness of the proposed algorithm, a proof-of-concept device was constructed using a commercially available circumaural hearing protector (Optime 98, Peltor, St. Paul, MN) to which were attached miniature loudspeakers (HD600, Sennheiser Electronic Corporation, Old Lyme, CT), and electret microphones (WM-61A, Panasonic, Newark, NJ). Signals to and from the electroacoustic components were conditioned using a custom-built analog front-end and sampled using analog-to-digital and digital-to-analog converters (ADS8361 and DAC8534, Texas Instruments, Dallas, TX). The algorithm was executed on a TMS320C6713B (Texas Instruments) digital signal processor (DSP) using a mixture of C and assembly programming languages. The device could be operated either as a traditional passive circumaural HPD, or as an electronic HPD with the warning signal detection algorithm. A computer model of the hearing protector was also generated that included the transfer functions from E to I and from L to I. The system identification methods used to obtain these models is described elsewhere (Bernstein et al., 2013). For testing purposes, an industry standardized alarm signal was generated to mix with the environmental noise source (ISO 7731, 2003). The alarm consisted of five pulses of a 1 kHz tone pulsed at a rate of 1 s with a 0.5 s on-period and a 0.5 s offperiod followed by a quiet interval. These parameters were based on information
J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection EL31
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
obtained from a fork lift truck and other back-up alarms. The environmental noise was recorded at the commander’s position of a Leopard 1 tank (Steeneken and Geurtsen, 1988). Human subject testing of the proposed algorithm implemented in the proof-of-concept device was conducted in an anechoic chamber using four loudspeaker towers (SRX715F and SRX718S, JBL, Stamford, CT) arranged in a distorted square to produce a pseudo-diffuse sound field in the horizontal plane at the center-head position. An alarm speaker (SRX715F) was placed approximately 1 m in front of the subject. Psychoacoustic testing used a “yes/no” forced choice paradigm to find the threshold of detection where a three-down/one-up method adapted the alarm signal SPL from an initial 10 dB signal-to-noise ratio (SNR) until six reversals were registered (Levitt, 1971). The performance of the device was evaluated by six subjects with normal hearing (4 male, 2 female). All gave their informed consent to participate in the study, which was conducted according to the provisions of the University of Connecticut Health Center’s Institutional Review Board. 3. Results Figure 2 shows the results of a computer simulation of the proposed method illustrating several of the signals generated by the alarm detection algorithm for various different alarm SNRs. The left hand columns demonstrate the performance for an SNR of þ10 dB. At this SNR, the alarm is visible in the signal recorded at the surface of the ear cup, e(n). The bandpass filtered signal, abp(n), demonstrates even better SNR once the out-of-band frequency components are removed, and provides a good illustration of the alarm signal waveform. In this application, the bandpass filter Hebp(z) was a sixth order Butterworth infinite impulse response filter with Q ¼ 10 and center frequency of 1 kHz. The low-pass filtered envelope signal, aen(n), shows the on and off cycle times of the alarm signal with very little noise. The envelop filter used in the simulation was a second order Butterworth lowpass filter with a cutoff frequency of 50 Hz. The detection threshold D was set to 200 (the dashed line of Fig. 2) and the signal d(j) shows that this threshold is exceeded in response to the alarm and drops quickly during the quiet periods. The threshold signal, T(j), shows that the alarm signal is accurately detected when it is enabled by d(j) at this SNR. The counter included in Eq. (3) eliminates any rapid pulsing of T(j) for a period after a detection, as d(j) periodically drops below the threshold value even during the detection period. Note that a
Fig. 2. Various signal waveforms used in the detection algorithm at different signal-to-noise ratios. Alarm signal is present from 0–4.5 s and 8–12.5 s for each SNR condition.
EL32 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
latency of approximately 1 s is seen from the time of the start of the first alarm pulse in e(n) to the time of the rising edge of T(j). This latency is a direct result of the selection of the coefficients of the basis vector, s. In this simulation, s was selected to include two 1 s on-off alarm pulses. A longer latency of approximately 2 s is seen from the end of the last alarm pulse in e(n) to the falling edge of T(j). This delay is due to setting M ¼ 100 for this simulation, as well as to the length of s. The center column of Fig. 2 shows the same signals at an SNR of 10 dB. Under this condition, the alarm signal is no longer distinguishable in e(n), but is still visible in the filtered signal abp(n). The envelope signal still shows the on-off periods of the alarm pulses, but the higher environmental noise SPL is more visible as noise than in the previous case. Comparing the aen(n) waveforms for the þ10 and 10 dB cases demonstrates the need for the vector normalization of Eq. (1), as the different SPLs lead to peak values of either 4 or 0.5, respectively. Despite the reduction in the SNR to 10 dB, the alarm signal is still detected by the correlator, and the signals d(j) and T(j) do not show any significant changes from the previous condition. The right column of Fig. 2 illustrates the detection limit of the algorithm at 30 dB SNR, for this noise source. The alarm signal is no longer detectable in the bandpass signal abp(n) and the envelope signal aen(n) does not show the on-off pulses of the alarm. The detection signal d(j) crosses the threshold once, but consistent detection of the alarm does not occur. For lower SNRs, the system does not detect any alarms for the entire duration of the source recording (approximately 2 min). Once the algorithm was validated in simulation, it was transferred to the proofof-concept device to evaluate its effect on the threshold needed by human subjects to detect the alarm signal in noise. In order to avoid subjects incurring hearing loss by poorly fitting, or removing, the HPD, the environmental noise was set to 85 dBA (re 2 105 Pa). For the environmental noise used in this study, the traditional passive attenuation of the circumaural HPD resulted in an average threshold of 60 (4.2) dBA [mean, and (standard deviation)] for detecting the warning sound. Subjects detected the warning sound at a threshold of 53 (4.6) dBA when wearing the proof-of-concept device with the algorithm operating. This represents a 7 dB improvement in the alarm detection threshold, and is statistically significant (p ¼ 0.037, two-sided t-test). 4. Discussion With the environmental noise SPL set to 85 dBA, the threshold for alarm signal detection by subjects when the algorithm was operating corresponded approximately to a 30 dB SNR. This value closely agrees with that obtained from computer simulation of the detection algorithm, and suggests that the subjects were relying on the algorithm to detect the alarm signal. Note that the alarm sounds in Fig. 2 have been plotted with the same time reference for each SNR, and so it is known that alarms occurred at times from approximately 0–4.5 s and from 8–12.5 s. Inspection of the response at 30 dB SNR reveals that the algorithm responded at about 11.5 s [where T(j) ¼ 1], that is, during an alarm, indicating it was not generating a false alarm. The similarity between the detection thresholds obtained in simulation and by subjects thus further suggests that the algorithm generates few false alarms. The influence on the reaction time to the alarm, however, remains to be assessed. In this algorithm, the alarm signal bypassed the attenuation of the hearing protector while the attenuation of the out-of-band environmental noise was maintained. This approach has the advantage of preserving some of the lateralization cues of the alarm, particularly the interaural time differences needed for left-right localization. However, bypassing the attenuation of the HPD potentially leads to high SNRs of the alarm signal with respect to the attenuated noise underneath the ear cup. This is because, in certain environments, the SPL of the alarm has been set with the expectation that it would be attenuated by hearing protection. Research has shown that the preferred auditory alarm SNR is between þ15 and þ25 dB, and values above this risk the device being turned off by users for being excessively loud (Edworthy and Hellier, 2000). As a result, rather than bypassing
J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection EL33
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
the passive attenuation of the ear cup, an alternative design could present the alarm signal within the preferred SNR range based on the SPL of the environmental noise underneath the ear cup. If signal detection is more important than relaying localization cues to the user, then the algorithm could be modified to reproduce a synthesized alarm signal upon its detection. A synthesized signal will maintain the same acoustic characteristics underneath the ear cup regardless of the outside alarm SNR, but the directionality of the alarm signal would be lost. This approach would be acceptable in environments where the presence of the alarm is more important than its direction, such as a fire alarm, but would not be appropriate for situations where localization is important, such as identifying the origin of a vehicle back-up alarm. The ability to provide directional information distinguishes the method described here from alternative alarm modalities, such as a radio link. Training the algorithm to detect multiple warning sounds occurring in an environment should, for example, enable workers to respond to a range of hazards in the workplace. As presently conceived, the algorithm could readily be integrated into existing digital electronic HPDs. 5. Conclusion This letter has described a method for improving the detection of an auditory warning alarm for persons wearing an HPD. The algorithm is designed to search for the envelope of a pre-selected alarm signal, and uses a normalized correlation detection threshold for identifying when the alarm is present in the external environment. The detected alarm signal is then allowed to bypass the passive attenuation of the HPD, while maintaining the attenuation of environmental noise. Computer simulation demonstrates the detection capability of the algorithm for a wide variety of SNRs. Implementation of the algorithm in a concept device worn by human subjects has shown that the method provides an improvement in the alarm detection threshold over a traditional HPD, and thus is expected to aid situational awareness in environments where hearing protectors are worn. Acknowledgments This work was supported by the National Institute for Occupational Safety and Health under research Grant No. R01 OH0086690-05. References and links Alali, K., and Casali, J. (2011). “The challenge of localizing vehicle backup alarms: Effects of passive and electronic hearing protectors, ambient noise level, and backup alarm spectral content,” Noise Health 13(51), 99–112. Bernstein, E. R., Brammer, A. J., and Yu, G. (2013). “Improving speech intelligibility in communication headsets: Simulation of adaptive subband processing for speech in noise,” Int. J. Ind. Ergon. 43(6), 526–535. Carbonneau, M., Lezzoum, N., Voix, J., and Gagnon, G. (2013) “Detection of alarms and warning signals on an digital in-ear device,” Int. J. Ind. Ergon. 43(6), 503–511. Casali, J., Robinson, G., Dabney, E., and Gauger, D. (2004). “Effect of electronic ANR and conventional hearing protectors on vehicle backup alarm detection in noise,” Hum. Factors 46(1), 1–10. Edworthy, J., and Hellier, E. (2000). “Auditory warning in noisy environments,” Noise Health 2(6), 27–39. ISO (2003). 7731, Ergonomics—Danger Signals for Public and Work Areas—Auditory Danger Signals (International Organization for Standards, Geneva), pp. 5–7. Levitt, H. (1971). “Transformed up-down methods in psychoacoustics,” J. Acoust. Soc. Am. 49(2B), 467–477. Lutfi, R., and Heo, I. (2012). “Automated detection of alarm sounds,” J. Acoust. Soc. Am. 132(2), EL125–EL128. Steeneken, H., and Geurtsen, F. (1988). “Description of the RSG-10 Noise Database,” TNO Institute for Perception, Report No. 1988-3, pp. 1–13.
EL34 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
Augmented warning sound detection for hearing protectors Eric R. Bernstein, Anthony J. Brammer, and Gongqiang Yu Ergonomic Technology Center, University of Connecticut Health Center, 263 Farmington Avenue, Farmington, Connecticut 06030-2017 [email protected], [email protected], [email protected]
Abstract: Perception of warning sounds, such as vehicle backup alarms, is reduced when hearing protection devices (HPDs) are worn. A crosscorrelation approach is employed to detect a pre-selected warning sound and enable it to bypass the attenuation of the HPD while still attenuating the environmental noise. Computer simulation shows that the algorithm can detect the specified alarm at signal-to-environmental-noise ratios as low as 30 dB. Human subject testing of the algorithm, implemented on a modified commercial HPD, confirms the minimum detection threshold obtained in simulation, and demonstrates a 7 dB improvement in detection threshold compared with the unmodified HPD. C 2013 Acoustical Society of America V
PACS numbers: 43.66.Vt, 43.60.Bf, 43.60.Mn [QJF] Date Received: October 1, 2013 Date Accepted: November 14, 2013
1. Introduction In high sound pressure level (SPL) environments, the perception of alarm signals is reduced when hearing protection devices (HPDs) are worn (Edworthy and Hellier, 2000). Two considerations are of importance in these situations: alarm detection and alarm localization (Casali et al., 2004; Alali and Casali, 2011). This paper explores a method for augmenting alarm signal detection when using HPDs in noisy environments. The proposed algorithm searches for a pre-selected alarm signal in the environment and, if found, allows that alarm signal to bypass the passive attenuation of the HPD while maintaining attenuation of the environmental noise. The normalized correlation-based approach used in this detection algorithm is similar to methods proposed by Lutfi and Heo (2012) and Carbonneau et al. (2013). However, these methods rely on the periodicity of the autocorrelation function for detecting a general alarm signal, rather than optimizing the detection of a specific alarm using the cross-correlation with the expected signal. The performance of the proposed algorithm has been evaluated by computer simulation as well as when applied to a modified commercial HPD worn by human subjects in environmental noise. 2. Methods Figure 1 shows a block diagram of the proposed alarm signal detection algorithm integrated into a circumaural hearing protector. The system uses an external microphone, E, mounted on the surface of the ear cup to detect the unwanted environmental noise mixed with the alarm sounds. The microphone signal, e(n), is passed through a filter, Hebp(z), to remove unnecessary frequency components that are outside the expected range of the tonal warning signal. For such warning signals, Hebp(z) would take the form of a bandpass filter with the passband centered around the expected alarm signal frequency. The absolute value of the resulting bandpassed alarm signal abp(n) is passed through a lowpass envelop filter, Hen(z), to generate aen(n), which is then downsampled by a factor of J. For this study, the sampling rate from the external microphone was set to 24 kHz and J was set to 96, resulting in a 250 Hz sampling rate for the alarm detection algorithm. The N most recent output samples of the downsampler are then
J. Acoust. Soc. Am. 135 (1), January 2014
C 2013 Acoustical Society of America V
EL29
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
Fig. 1. Block diagram of proposed alarm signal detection algorithm implemented in a circumaural hearing protector.
stored in the vector r(j) for use in the correlation stage of the algorithm. Before using r(j) for alarm detection, the vector must be normalized to account for the wide dynamic range of sounds in the real world. The normalized vector rn(j) is given by rn ðjÞ ¼
rðjÞ r ðjÞ ; rr ðjÞ
(1)
where r ðjÞ and rr ðjÞ are the mean and standard deviation of r(j), respectively. The normalized signal vector is then cross-correlated with a length N basis vector, s, to generate the detection signal, d(j), according to N 1 X dðjÞ ¼ sðiÞrn ðj iÞ:
(2)
i¼0
The coefficients of the vector s are tuned to the characteristics of the alarm signal that the algorithm is designed to detect. The s vector coefficients are determined off-line and are given by the vector rn(j) from Eq. (1) under the condition when the alarm signal is present with no environmental noise. The detection signal, d(j), is then used in the “Threshold” block to generate a binary signal, T(j), that takes a zero value if the alarm is not present and a unity value if the alarm signal is present. T(j) is thus found according to 1 if dðjÞ D or m > 0; TðjÞ ¼ (3) 0 else; where D is the threshold parameter and m is a counter decremented at each algorithm iteration to ensure T(j) maintains a value of unity for some time after a threshold has been exceeded. For this purpose, m is set to the value of a count parameter, M, when an “alarm” has been detected. This enables a suitable gating signal to be generated for gain control. Once a signal has been detected, the system must alert the user to the presence of the alarm. In this paper, the alarm signal is presented to the user at the SPL that occurs in the environment, and thus the alarm signal (but not the environmental noise) should bypass the attenuation of the passive ear cup. To accomplish this, the bandpass
EL30 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
filtered signal abp(n) is passed through a variable gain block G(j) to generate the output signal y(n) for loudspeaker, L, located underneath the ear cup. In this way, the alarm noise from outside the ear cup is reproduced underneath the ear cup, while the environmental noise outside the narrow passband of abp(n) remains attenuated. To establish the appropriate gain, it is first necessary to determine the environmental noise SPL underneath the ear cup. For this purpose, the sound generated by the loudspeaker must be removed from that sensed by the internal microphone, I[i(n)]. An estimate of the algorithm-generated warning sound to subtract from i(n) is obtained by passing y(n) through a transfer function model from L to I, which represents the electro-acoustic transfer function of the loudspeaker and microphone as well as the sound propagation from L to I. The resulting signal y0 (n) is subtracted from i(n) to yield i0 (n). The signal i0 (n) is then passed through a bandpass filter, Hibp(z), to generate ibp(n), which is used to measure the SPL of the noise underneath the ear cup. The SPL estimates of abp(n) and ibp(n), namely, A(n) and I(n), respectively, are found according to AðnÞ ¼ b Aðn 1Þ þ ð1 bÞjabp ðnÞj2
(4)
I ðnÞ ¼ aI ðn 1Þ þ ð1 aÞjibp ðnÞj2 ;
(5)
and
where a and b are exponential forgetting parameters usually taking values close to unity. If an alarm signal is not detected, G(j) takes a value of zero, thereby disabling electronic transmission of sound from the environment to within the ear cup. If an alarm signal is detected, the gain should be such that the alarm signal underneath the ear cup has the same SPL as outside, and thus the gain must take into account the attenuation of the earmuff. The gain G(j) is then 8 > 0 if TðjÞ ¼ 0; > < sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (6) GðjÞ ¼ AðjÞ I ðjÞ > if TðjÞ ¼ 1: > : AðjÞ þ c Using the gain selected, the signal underneath the ear cup is set so that the summation of the alarm sound generated by the loudspeaker and alarm sound as attenuated by the HPD should maintain the same SPL as the alarm signal outside the ear cup. To evaluate the effectiveness of the proposed algorithm, a proof-of-concept device was constructed using a commercially available circumaural hearing protector (Optime 98, Peltor, St. Paul, MN) to which were attached miniature loudspeakers (HD600, Sennheiser Electronic Corporation, Old Lyme, CT), and electret microphones (WM-61A, Panasonic, Newark, NJ). Signals to and from the electroacoustic components were conditioned using a custom-built analog front-end and sampled using analog-to-digital and digital-to-analog converters (ADS8361 and DAC8534, Texas Instruments, Dallas, TX). The algorithm was executed on a TMS320C6713B (Texas Instruments) digital signal processor (DSP) using a mixture of C and assembly programming languages. The device could be operated either as a traditional passive circumaural HPD, or as an electronic HPD with the warning signal detection algorithm. A computer model of the hearing protector was also generated that included the transfer functions from E to I and from L to I. The system identification methods used to obtain these models is described elsewhere (Bernstein et al., 2013). For testing purposes, an industry standardized alarm signal was generated to mix with the environmental noise source (ISO 7731, 2003). The alarm consisted of five pulses of a 1 kHz tone pulsed at a rate of 1 s with a 0.5 s on-period and a 0.5 s offperiod followed by a quiet interval. These parameters were based on information
J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection EL31
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
obtained from a fork lift truck and other back-up alarms. The environmental noise was recorded at the commander’s position of a Leopard 1 tank (Steeneken and Geurtsen, 1988). Human subject testing of the proposed algorithm implemented in the proof-of-concept device was conducted in an anechoic chamber using four loudspeaker towers (SRX715F and SRX718S, JBL, Stamford, CT) arranged in a distorted square to produce a pseudo-diffuse sound field in the horizontal plane at the center-head position. An alarm speaker (SRX715F) was placed approximately 1 m in front of the subject. Psychoacoustic testing used a “yes/no” forced choice paradigm to find the threshold of detection where a three-down/one-up method adapted the alarm signal SPL from an initial 10 dB signal-to-noise ratio (SNR) until six reversals were registered (Levitt, 1971). The performance of the device was evaluated by six subjects with normal hearing (4 male, 2 female). All gave their informed consent to participate in the study, which was conducted according to the provisions of the University of Connecticut Health Center’s Institutional Review Board. 3. Results Figure 2 shows the results of a computer simulation of the proposed method illustrating several of the signals generated by the alarm detection algorithm for various different alarm SNRs. The left hand columns demonstrate the performance for an SNR of þ10 dB. At this SNR, the alarm is visible in the signal recorded at the surface of the ear cup, e(n). The bandpass filtered signal, abp(n), demonstrates even better SNR once the out-of-band frequency components are removed, and provides a good illustration of the alarm signal waveform. In this application, the bandpass filter Hebp(z) was a sixth order Butterworth infinite impulse response filter with Q ¼ 10 and center frequency of 1 kHz. The low-pass filtered envelope signal, aen(n), shows the on and off cycle times of the alarm signal with very little noise. The envelop filter used in the simulation was a second order Butterworth lowpass filter with a cutoff frequency of 50 Hz. The detection threshold D was set to 200 (the dashed line of Fig. 2) and the signal d(j) shows that this threshold is exceeded in response to the alarm and drops quickly during the quiet periods. The threshold signal, T(j), shows that the alarm signal is accurately detected when it is enabled by d(j) at this SNR. The counter included in Eq. (3) eliminates any rapid pulsing of T(j) for a period after a detection, as d(j) periodically drops below the threshold value even during the detection period. Note that a
Fig. 2. Various signal waveforms used in the detection algorithm at different signal-to-noise ratios. Alarm signal is present from 0–4.5 s and 8–12.5 s for each SNR condition.
EL32 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
latency of approximately 1 s is seen from the time of the start of the first alarm pulse in e(n) to the time of the rising edge of T(j). This latency is a direct result of the selection of the coefficients of the basis vector, s. In this simulation, s was selected to include two 1 s on-off alarm pulses. A longer latency of approximately 2 s is seen from the end of the last alarm pulse in e(n) to the falling edge of T(j). This delay is due to setting M ¼ 100 for this simulation, as well as to the length of s. The center column of Fig. 2 shows the same signals at an SNR of 10 dB. Under this condition, the alarm signal is no longer distinguishable in e(n), but is still visible in the filtered signal abp(n). The envelope signal still shows the on-off periods of the alarm pulses, but the higher environmental noise SPL is more visible as noise than in the previous case. Comparing the aen(n) waveforms for the þ10 and 10 dB cases demonstrates the need for the vector normalization of Eq. (1), as the different SPLs lead to peak values of either 4 or 0.5, respectively. Despite the reduction in the SNR to 10 dB, the alarm signal is still detected by the correlator, and the signals d(j) and T(j) do not show any significant changes from the previous condition. The right column of Fig. 2 illustrates the detection limit of the algorithm at 30 dB SNR, for this noise source. The alarm signal is no longer detectable in the bandpass signal abp(n) and the envelope signal aen(n) does not show the on-off pulses of the alarm. The detection signal d(j) crosses the threshold once, but consistent detection of the alarm does not occur. For lower SNRs, the system does not detect any alarms for the entire duration of the source recording (approximately 2 min). Once the algorithm was validated in simulation, it was transferred to the proofof-concept device to evaluate its effect on the threshold needed by human subjects to detect the alarm signal in noise. In order to avoid subjects incurring hearing loss by poorly fitting, or removing, the HPD, the environmental noise was set to 85 dBA (re 2 105 Pa). For the environmental noise used in this study, the traditional passive attenuation of the circumaural HPD resulted in an average threshold of 60 (4.2) dBA [mean, and (standard deviation)] for detecting the warning sound. Subjects detected the warning sound at a threshold of 53 (4.6) dBA when wearing the proof-of-concept device with the algorithm operating. This represents a 7 dB improvement in the alarm detection threshold, and is statistically significant (p ¼ 0.037, two-sided t-test). 4. Discussion With the environmental noise SPL set to 85 dBA, the threshold for alarm signal detection by subjects when the algorithm was operating corresponded approximately to a 30 dB SNR. This value closely agrees with that obtained from computer simulation of the detection algorithm, and suggests that the subjects were relying on the algorithm to detect the alarm signal. Note that the alarm sounds in Fig. 2 have been plotted with the same time reference for each SNR, and so it is known that alarms occurred at times from approximately 0–4.5 s and from 8–12.5 s. Inspection of the response at 30 dB SNR reveals that the algorithm responded at about 11.5 s [where T(j) ¼ 1], that is, during an alarm, indicating it was not generating a false alarm. The similarity between the detection thresholds obtained in simulation and by subjects thus further suggests that the algorithm generates few false alarms. The influence on the reaction time to the alarm, however, remains to be assessed. In this algorithm, the alarm signal bypassed the attenuation of the hearing protector while the attenuation of the out-of-band environmental noise was maintained. This approach has the advantage of preserving some of the lateralization cues of the alarm, particularly the interaural time differences needed for left-right localization. However, bypassing the attenuation of the HPD potentially leads to high SNRs of the alarm signal with respect to the attenuated noise underneath the ear cup. This is because, in certain environments, the SPL of the alarm has been set with the expectation that it would be attenuated by hearing protection. Research has shown that the preferred auditory alarm SNR is between þ15 and þ25 dB, and values above this risk the device being turned off by users for being excessively loud (Edworthy and Hellier, 2000). As a result, rather than bypassing
J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection EL33
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
Bernstein et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4835135]
Published Online 11 December 2013
the passive attenuation of the ear cup, an alternative design could present the alarm signal within the preferred SNR range based on the SPL of the environmental noise underneath the ear cup. If signal detection is more important than relaying localization cues to the user, then the algorithm could be modified to reproduce a synthesized alarm signal upon its detection. A synthesized signal will maintain the same acoustic characteristics underneath the ear cup regardless of the outside alarm SNR, but the directionality of the alarm signal would be lost. This approach would be acceptable in environments where the presence of the alarm is more important than its direction, such as a fire alarm, but would not be appropriate for situations where localization is important, such as identifying the origin of a vehicle back-up alarm. The ability to provide directional information distinguishes the method described here from alternative alarm modalities, such as a radio link. Training the algorithm to detect multiple warning sounds occurring in an environment should, for example, enable workers to respond to a range of hazards in the workplace. As presently conceived, the algorithm could readily be integrated into existing digital electronic HPDs. 5. Conclusion This letter has described a method for improving the detection of an auditory warning alarm for persons wearing an HPD. The algorithm is designed to search for the envelope of a pre-selected alarm signal, and uses a normalized correlation detection threshold for identifying when the alarm is present in the external environment. The detected alarm signal is then allowed to bypass the passive attenuation of the HPD, while maintaining the attenuation of environmental noise. Computer simulation demonstrates the detection capability of the algorithm for a wide variety of SNRs. Implementation of the algorithm in a concept device worn by human subjects has shown that the method provides an improvement in the alarm detection threshold over a traditional HPD, and thus is expected to aid situational awareness in environments where hearing protectors are worn. Acknowledgments This work was supported by the National Institute for Occupational Safety and Health under research Grant No. R01 OH0086690-05. References and links Alali, K., and Casali, J. (2011). “The challenge of localizing vehicle backup alarms: Effects of passive and electronic hearing protectors, ambient noise level, and backup alarm spectral content,” Noise Health 13(51), 99–112. Bernstein, E. R., Brammer, A. J., and Yu, G. (2013). “Improving speech intelligibility in communication headsets: Simulation of adaptive subband processing for speech in noise,” Int. J. Ind. Ergon. 43(6), 526–535. Carbonneau, M., Lezzoum, N., Voix, J., and Gagnon, G. (2013) “Detection of alarms and warning signals on an digital in-ear device,” Int. J. Ind. Ergon. 43(6), 503–511. Casali, J., Robinson, G., Dabney, E., and Gauger, D. (2004). “Effect of electronic ANR and conventional hearing protectors on vehicle backup alarm detection in noise,” Hum. Factors 46(1), 1–10. Edworthy, J., and Hellier, E. (2000). “Auditory warning in noisy environments,” Noise Health 2(6), 27–39. ISO (2003). 7731, Ergonomics—Danger Signals for Public and Work Areas—Auditory Danger Signals (International Organization for Standards, Geneva), pp. 5–7. Levitt, H. (1971). “Transformed up-down methods in psychoacoustics,” J. Acoust. Soc. Am. 49(2B), 467–477. Lutfi, R., and Heo, I. (2012). “Automated detection of alarm sounds,” J. Acoust. Soc. Am. 132(2), EL125–EL128. Steeneken, H., and Geurtsen, F. (1988). “Description of the RSG-10 Noise Database,” TNO Institute for Perception, Report No. 1988-3, pp. 1–13.
EL34 J. Acoust. Soc. Am. 135 (1), January 2014
Bernstein et al.: Augmented warning sound detection
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 134.117.10.200 On: Wed, 17 Jun 2015 18:45:37
