A vocal repertoire of Asian elephants (Elephas maximus) and comparison of classification methods Sharon S. Glaeser1, Holger Klinck2, David K. Mellinger2, Yao Ren3, Patrick J. Clemins4, Michael T. Johnson3, Randy Zelick1 (1) Department of Biology, Portland State University (2) Cooperative Institute of Marine Resources, Oregon State University and National Oceanic and Atmospheric Administration (NOAA) Pacific Marine Environmental Laboratory (3) Department of Electrical and Computer Engineering, Marquette University (4) AAAS Science and Technology Policy Fellow, National Science Foundation Steps in conducting this research • Conducted a bibliographic research, talked to elephant experts • Developed research questions • Conducted reconnaissance observations, discussed data collection options with elephant management • Constructed an ethogram of elephant call definitions • Designed the study • Type of study: observational (not experimental) • Recording media: audio, video • Sampling rule: behavior sampling • Recording rule: continuous recording • Collected data • Analyzed data • Interpreted results and ask new questions Original research questions v. final goals • Original research ideas included: • Evaluation of calls as communication signals, but this requires examination of the response by conspecifics. • Evaluation of vocalizations produced by males in musth. • Original research questions required well-defined call types: • How are calls associated with social context, behavior, perceived emotional state, reproductive state, or an individual? • How is variation of acoustic parameters within a call type associated with social context, perceived emotional state, reproductive state, individuality, or size? • Bibliographic research showed calls were not well defined by acoustic parameters; rather most were defined by context. • Reconnaissance observations showed that males in this group were not producing a vocalization unique to musth. Contribution to basic science • Define an acoustic repertoire of Asian elephants based on acoustic parameters Goals Ecology Acoustics Data Collection • Validate structural distinction among call types using computerized methods • Investigate how the repertoire is used by groups and individuals • Provide a basis for comparing acoustic communication among elephant species and populations Analysis Sharon Stanton Distribution Summary Ecology Goals Ecology Acoustics Data Collection Analysis • We try to offer our captive elephants an environment that challenges them to utilize behaviors they would exhibit in the wild. • Acoustic signals serve social functions, so it is important to take into account the differences in social structure of groups under study and how these differences may impact their communication. Distribution Summary • And...we are encouraged when we discover our group in Oregon utilizes an acoustic repertoire as diverse as is found in more natural social networks in range countries (Glaeser, 2009; Nair et al., 2009; de Silva, 2010). Social structure of wild elephants • Complex, fluid society Goals Ecology • Multigenerational matriarchal herds of females and young with bulls on the periphery Acoustics • Mature bulls live in loose association with other bulls, except in musth when they travel alone and seek out females Data Collection • Evidence that bull society may be more structured than previously thought Analysis Distribution Summary – Young bulls seek company of mature bulls • Society is multi-tiered, family units form bond groups or larger clans that share home range – Asian elephant society may demonstrate less social connectivity at the population level, and may be less influenced by seasonal differences in ecological conditions Captive elephants (N. America & Europe) • Group structure – Females Goals Ecology Acoustics Data Collection Analysis Distribution Summary • Natal groups + transfers – Males • Adult bulls separated (few exceptions) • Adult bulls sometimes in with females • Social introductions – Male/female introductions for breeding and socialization – Female/female introductions with change in group structure • Temporary separations – Breeding – Husbandry Domesticated elephants (SE Asia) • Group structure – Females Goals • Natal groups + transfers – Males Ecology Acoustics Data Collection Analysis Distribution Summary • Adult bulls work in close proximity but separated on night chains • Adult bulls separated during musth (or maybe if known to kill) – Mahout (handler) with elephant most of the time – Structure can be very fluid (animals bought/sold/leased) • Social introductions – Common with fluid structure – Management varies – May be restrained • Temporary separations – Work, loans – husbandry Communication modalities • Long-distance: low frequency vocalizations, chemical signals [1][2] Goals Ecology • Short-distance: visual, tactile, auditory, and chemical signals [3] Acoustics • Seismic detection and discrimination of vocalizations [4][5] Data Collection Analysis Distribution Summary [1]Payne et al. 1986 [2]Poole 1999 [3]Schulte & Rasmussen 1999 [4]O’Connel-Rodwell et al. 2006, [5]O’Connel-Rodwell et al. 2007 Process of studying sound Record sound (microphone responds to sound pressure) Create visual representation of sound Measure sound characteristics (similar to measuring morphological characteristics) How do we visualize sound? Spectrogram: Shows frequency change with time Shows relative intensity with color Waveform: Shows amplitude change with time Produced by measuring device (mic) Power spectrum: Shows relative power of frequencies Power is averaged across signal What did we know about acoustic communication of Asian elephants at start of study? Goals African elephant Asian elephant Calls defined 25+ [1] (by context) many rumbles 6 calls [1][2][10] (by context) Frequency range 5Hz to 9kHz [3] 14 Hz [4] to ? Musth call musth rumble(M) [5] ?? Individual recognition contact call (F) [6][7] distinct rumble (F) [8] ?? Ecology Acoustics Data Collection Analysis Emotional state of caller unstable pitch (F) [8] ?? Increased motivation, intention modifications of basic calls [2] Distribution Summary call combination in calves [9] [1]Olson 2005 (Elephant Husbandry Guide) [2]McKay, 1973 [3]Poole in prep [4]Payne et al. 1986 [5]Poole et al. 1988 [6]McComb et al. 2003 [7]Langbauer et al. 1991 [8]Soltis et al. 2005 [9]Stoeger-Harwath and Kratochvil 2007 [10]Nair et al., 2009 What did we discover with this study? Goals African elephant Asian elephant Calls defined 25+ [1] (by context) many rumbles 6 basic call types 5 combinations calls Frequency range 5Hz to 9kHz [3] 14 Hz to 18 kHz Musth call musth rumble(M) [5] ?? Individual recognition contact call (F) [6][7] distinct rumble (F) [8] ?? Ecology Acoustics Data Collection Analysis Emotional state of caller unstable pitch (F) [8] ?? Increased motivation, intention modifications of basic calls [2] Distribution Summary call combination in calves [9] [1]Olson 2005 (Elephant Husbandry Guide) [2]McKay, 1973 [3]Poole in prep [4]Payne et al. 1986 [5]Poole et al. 1988 [6]McComb et al. 2003 [7]Langbauer et al. 1991 [8]Soltis et al. 2005 [9]Stoeger-Harwath and Kratochvil 2007 Study sites • Oregon Zoo (Portland, OR) – Urban setting Goals Ecology Acoustics Data Collection Analysis Distribution Summary • Anthropogenic noise: highway traffic, air traffic, water, electricity, hydraulics, machinery, train, people • Royal Elephant Kraal (Ayutthaya, central Thailand) – Working village – Urban setting • Anthropogenic noise: road traffic, water, machinery, people • Elephant Nature Park (Mae Taeng Valley, northern Thailand) – Tourist park – Rural setting • Anthropogenic noise: construction, sporadic road traffic, people Study subjects – Oregon Zoo Pet(f) matriarch, at age 51 Packy (m) Tusko (m) Rama (m) Sung-Surin (f) Rose Tu (f) Chendra (f) All adults in the study Study subjects – Thailand • Captive elephants • Royal Elephant Kraal ~80 elephants • males and females of all ages • geriatric animals • 10 -12 calves • Elephant Nature Park ~30 elephants • • • • 3 adult bulls 2 sub-adult males 2 calves females of varying ages Recording methods • Data collection at Oregon Zoo Goals Ecology Acoustics Data Collection Analysis Distribution Summary – February 2005 to March 2009 – Recorded continuously (typically 1 hour) Context: social introductions, breeding, temporary separations, arrival of bull, euthanasia of matriarch, routine husbandry 56 hours of data (usable data) • Data collection in Thailand – – November & December 2009 Recorded continuously (typically 1 hour) Context: morning release to pasture, painting, night feedings, routine 6 hours of data • Recording equipment – – – 32 to 44 kHz sampling rate M-Audio MicroTrack II (20 Hz to 20 kHz, tested to 5 Hz) Digital video (audio extraction, 20 Hz to 20 kHz) Oregon zoo only • Edirol R-09 mp3 recorder (20 Hz to 40 KHz, recorded lower) • Racal V-Store recorder, Bruel & Kjaer 4145 microphone (2.6 Hz to 10 kHz) Caller identification • Localizing calls Goals Ecology Acoustics Data Collection – Live observations only – Depends on distance between subjects • Visual cues for sound production – Air blast from trunk or mouth (e.g., sand, condensation) – Mouth wide open – Depressed cheeks w/ squeal and squeak Analysis Distribution quiet: cheeks relaxed Summary squeak & squeal: cheeks depressed Analysis overview Categorized calls by listening to calls & viewing spectrograms (N=2791, interobserver reliability = 93%) Measured sound characteristics with acoustic software then ran statistics Used pattern recognition (speech recognition) to test categories basic call types Basic call types Bark Roar Rumble Bark Squeak Squeals Trumpet Combinations / Modifications • Repetition of basic call • Combination of basic types with continuous frequency contour Squeak train Squeal + Squeak Squeak + Bark Roar + Rumble Trumpet + Roar Analysis overview Categorized calls by listening to calls & viewing spectrograms (N=2791, interobserver reliability = 93%) Measured sound characteristics with acoustic software then ran statistics Used pattern recognition (speech recognition) to test categories basic call types Measuring acoustic parameters Steps for measuring 1. Goals selection box 2. Ecology Acoustics Data Collection 3. feature box 4. Analysis User draws selection box liberally around signal Osprey creates feature box encompassing strongest 90% of energy Osprey extracts 28 acoustic features within feature box Osprey extracts signal-to-noise ratio within selection box Distribution Summary • Low frequency calls with energy near 20 Hz – Verified fundamental frequency by measuring harmonic spacing – Drew selection box to match fundamental in order to determine frequency limit at low end Acoustic parameters (examples) Goals Ecology Acoustics Data Collection Analysis • – – Distribution • Summary Temporal features Duration (M5) = duration of 90% of the energy Median time (M7) = time at with 50% energy reached Frequency features – – – – – Bandwidth (M6) = frequency range of 90% of energy Median time (M11) = frequency at with 50% energy reached Frequency of peak intensity (M20) = frequency of peak power Cepstrum peak width (M25) = average width of peaks in power spectrum Upsweep Mean (M27) = direction of frequency change Analysis overview Categorized calls by listening to calls & viewing spectrograms (N=2791, interobserver reliability = 93%) Measured sound characteristics with acoustic software then ran statistics Used pattern recognition (speech recognition) to test categories basic call types Tested call categories using pattern recognition • Hidden Markov Model Goals Ecology – Raw sound files, no noise cancellation or noise filtering – N=1686 (basic call types, no exclusions) • Success rate Acoustics Success Rate Data Collection Analysis Distribution Summary Quality 1 Quality 2 Quality 3 Quality 4 Overall 85.0% 76.3% 78.7% 67.7% 72.6% Quality 1 = No overlapping sound / less degredation by noise Quality 2 = No overlapping sound / more degredation by noise Quality 3 = Overlapping sound / less degredation by noise Quality 4 = Overlapping sound / more degredation by noise • Confusion with – Rumbles/Roars (not easy to hear difference sometimes) – Squeals/Trumpets (easy to hear difference, but frequency is similar) – Squeals/Rumbles (easy to hear difference, but duration is similar) Application for pattern recognition model • Scenario: Recorded one group, then test on a geographically and socially disparate group Goals Ecology Acoustics Data Collection Analysis Distribution Summary • Trained with Oregon calls, then tested with Thailand calls • Success rate dropped to 40.8% • Lower success could be due to difference in repertoire usage or to variability within call types • Could investigate similarities and differences among groups Are the call types distinct? • Classification tree showed overall success of 87.8% – Potential for classifying new calls into defined types Goals Ecology • PCA showed some support for grouping call types, but with some overlap between call types Acoustics • HMM showed overall successful classification 72.6% Data Collection Analysis Distribution Summary Are the call types significantly different based on acoustic parameters? YES • Global and pair-wise Analysis of Similarity (ANOSIM) test showed significant difference among and between all pairs of call types Distribution of calls: by site Thailand (N=279) Oregon Zoo (N=2066) Combo 13% Goals Ecology Combo 3% Rumble 6% Roar 11% Bark 4% Trumpet 12% Rumble 18% Trumpet 37% Roar 12% Acoustics Data Collection Analysis Squeak 25% Squeak 3% Squeal 29% Squeal 23% Distribution Summary Bark 4% • Call repertoire is the same (except Squeak) • Call composition is different – Number of subjects different – Contexts are different N=2345 excludes: blows calls on command Distribution of calls: by individual Pet Goals Trumpet 0% Sung-Surin Rumble Roar Combo 1% 1% 3% Bark 0% Combo 12% Squeak 0% Trumpet 0% Rose Tu Combo 5% Rumble 1% Roar Bark 3% 3% Rumble 12% Roar 9% Trumpet 36% Ecology Squeal 27% Bark 17% Acoustics Squeak 54% Squeal 95% Data Collection Squeal 12% Chendra Analysis Distribution Combo 3% Squeak 9% Tusko Rumble 8% Rumble 0% Roar 6% Trumpet 31% Bark 11% Squeak 0% Squeal 0% Trumpet 2% Combo 0% Summary Squeal 0% Squeak 0% Bark 3% Roar 55% Blow -tonal 81% N=1017 Rate of calling: by social context • Oregon Zoo Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals • Rate of calling adjusted to 60 min 50 40 30 20 10 ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R hm en t En ri c hr ic In tro du ct io n af te rI nt ro du ct io n eu ni on R Distribution R Analysis ou ti n e 0 Context Summary N=2147 excludes: blows calls on command males by themselves Rate of calling: by social context Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals 50 40 30 20 10 Context Summary ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R hm en t En ri c hr ic In tro du ct io n af te rI nt ro du ct io n eu ni on R Distribution R Analysis ou ti n e 0 Rate of calling: by social context Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals 50 40 30 20 10 Context Summary ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R hm en t En ri c hr ic In tro du ct io n af te rI nt ro du ct io n eu ni on R Distribution R Analysis ou ti n e 0 Rate of calling: by social context Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals 50 40 30 20 10 Context Summary ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R hm en t En ri c hr ic In tro du ct io n af te rI nt ro du ct io n eu ni on R Distribution R Analysis ou ti n e 0 Rate of calling: by social context Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals 50 40 30 Surprisingly low call rates in some contexts 20 10 Context Summary ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R hm en t En ri c hr ic In tro du ct io n af te rI nt ro du ct io n eu ni on R Distribution R Analysis ou ti n e 0 Rate of calling: by social context Call Rate by Context 70 60 Ecology Acoustics Data Collection Mean Call Rate/Hour Goals 50 40 30 20 10 ea th D Tr an sf er ep ar at io n af te rS Se pa ra tio n eu ni on R eu ni on In tro du ct io n af te rI nt ro du ct io n hm en t En ri c hr ic R Distribution R Analysis ou ti n e 0 Context Summary Rate of calling is not a reliable indicator of arousal OR What we perceive as situations that warrant increased arousal do not for certain individuals or groups Summary of results • 6 basic call types that are significantly different – Bark, Roar, Rumble, Squeak, Squeal, Trumpet Goals Ecology Acoustics Data Collection Analysis Distribution Summary • 11 acoustic parameters differentiate the basic call types – Duration, Lower frequency, Upper frequency, Bandwidth, Median frequency, Frequency of overall intensity, Frequency modulation, Frequency modulation variation, Amplitude modulation, Amplitude modulation variation, Overall entropy • Recorded frequency range was 14 Hz to 18 KHz – Most calls are entirely audible – Frequency range for only 90% of signal energy 14 Hz to 9.3 kHz • Statistical analysis and pattern recognition model were consistent with categorization of calls by aural perception (what we hear) • Individuals use the repertoire differently, so we need to be cautious in assigning functional relevance Summary of results Example call description: Squeak Sounds like: Wet finger across rubber surface, baby alligator, audible pedestrian crossing signal Visual cues: Cheeks depressed, mouth open slightly or closed measurements of 90% of energy Lower frequency 194 – 883 Hz (median 495) Upper frequency 797 – 3768 Hz (median 1830) Bandwidth 301 – 3273 Hz (median 1318) Frequency of peak intensity 537– 2024 Hz (median 775) Duration 0.1 – 1.2 sec (median 0.2) Frequency modulation 1.1 – 70.5 Hz (median 4.9) Amplitude modulation 0.8 – 44.9 Hz (median 4.7) Overall Entropy 58 – 294 Hz (median 136) Future directions • Compare results with those from other study groups • Determine functional relevance of variability (conservatively) – Associate calls with individuals, social context, behavior, and perceived level of arousal Thank You Oregon Zoo elephant handlers: Jeb Barsh, Dimas Dominguez, Jeremy Kirby, Pat Flora, Bob Lee, Joe Sebastiani, Pam Starkey, April Yoder Oregon Zoo Interum Director: Mike Keele Oregon Zoo Conservation staff: Anne Warner, David Shepherdson, Karen Lewis Royal Elephant Kraal, Thailand Elephant Nature Park, Thailand Mandy L. Cook for teaching me how to measure sound Jean Lea Hoffheimer & Bobbi Estabrook for scoring calls Questions
© Copyright 2026 Paperzz