Separating overlapping click trains originating from multiple individuals in echolocation recordings

Recordings of the acoustic activity of free-swimming groups of echolocating dolphins increase the likelihood of collecting overlapping click trains, originating from multiple individuals, in the same set of data. In order to evaluate the click properties of each individual based on such recordings it is necessary to identify which clicks originate from which animal. This paper suggests a computationally efficient strategy to separate overlapping click trains originating from multiple free-swimming bottlenose dolphins, enabling echolocation analysis at an individual level on several animals. This technique is based on sequential matching of the frequency spectra of successive clicks. The clicks are grouped together as individual click trains if the correlation coefficients between clicks are higher than a pre-set threshold level. The robustness of the algorithm is tested by adding artificially generated white Gaussian noise and comparing the results with other comparable commonly used methods based on inter-click intervals, centroid frequencies, and amplitude levels. The described method is applicable to a variety of experimental and observational contexts, e.g., those regarding echolocation development of calves, the hypothesized acoustic "etiquette" among dolphins when investigating the same object, and the possible occurrence of eavesdropping in large dolphin pods.     

A generalized power-law detection algorithm for humpback whale vocalizations

Conventional detection of humpback vocalizations is often based on frequency summation of band-limited spectrograms under the assumption that energy (square of the Fourier amplitude) is the appropriate metric. Power-law detectors allow for a higher power of the Fourier amplitude, appropriate when the signal occupies a limited but unknown subset of these frequencies. Shipping noise is non-stationary and colored and problematic for many marine mammal detection algorithms. Modifications to the standard power-law form are introduced to minimize the effects of this noise. These same modifications also allow for a fixed detection threshold, applicable to broadly varying ocean acoustic environments. The detection algorithm is general enough to detect all types of humpback vocalizations. Tests presented in this paper show this algorithm matches human detection performance with an acceptably small probability of false alarms (P(FA) < 6%) for even the noisiest environments. The detector outperforms energy detection techniques, providing a probability of detection P(D) = 95% for P(FA) < 5% for three acoustic deployments, compared to P(FA) > 40% for two energy-based techniques. The generalized power-law detector also can be used for basic parameter estimation and can be adapted for other types of transient sounds.     

3D tracking of foraging belugas from their clicks: Experiment from a coastal hydrophone array

A simple passive acoustic monitoring (PAM) setup was used to localize and track beluga whales underwater in three dimensions (3D) in a fjord. In June 2009, beluga clicks were recorded from a cabled hydrophone array in a regularly frequented habitat in Eastern Canada. Beluga click energy was concentrated in the 30-50 kHz frequency band. The click trains detected on several hydrophones were localized from their time difference of arrivals. Cluster analysis linked localizations into tracks based on criteria of spatial and temporal proximity. At close ranges from the array, the localized click-train series allowed three-dimensional tracking of a beluga during its dive. Clicks within a train spanned a large range of durations, inter-click intervals, source levels and bandwidths. Buzzes sometimes terminated the trains. Repeated click packets were frequent. All click characteristics are consistent with oblique observations from the beam axis, and ordered variation of the source pattern during a train, likely resulting from a scan of angles from the beam axis, was observed before click trains indicated focusing of the echolocation clicks in one direction. The click-train series is interpreted as echolocation chasing for preys during a foraging dive. Results show that a simple PAM system can be configured to passively and effectively 3D track wild belugas and small odontocetes in their regularly frequented habitat.     

Real-time acoustic classification of sperm whale clicks and shipping impulses from deep-sea observatories

The automated real-time detection and classification of cetacean and anthropogenic sounds from deep-sea observatories can play a key role to study cetaceans in the field, to quantify the impact of anthropogenic sounds or to initiate mitigation measures during potentially harmful human activities. In the area of the NEMO-ONDE deep-sea observatory, sperm whales are often present together with heavy shipping. The spatial coincidence of both sound sources allows for the long term monitoring of their interaction. Some ships produce impulsive sounds and the automated separation of these impulses from sperm whale clicks is not a trivial task. As part of a detection, classification and localisation system for acoustic data from marine observatories, we present four modules performing the automated real-time classification of clicks from sperm whales and impulsive sounds produced by ships. First, two modules detect segments that contain impulsive sounds within a specifiable frequency band and return the impulses’ positions. Then, two modules classify the detected impulses as sperm whale clicks or ship impulses. Finally, at the level of 22 s segments, the classification outputs from individual impulses are combined into a decision on the presence of sperm whale clicks or ship impulses. The modules’ reliability was tested on data from the NEMO-ONDE observatory. Training and testing data were separated by more than 2 months, enabling to assess the consistency of the predictions over the long term. The automated separation between segments of the two classes was high with area under the ROC curve (AUC) values between 0.94 and 0.98.     

Masking-level differences for trains of clicks

Masking-level differences (MLDs) were measured for trains of 2000-Hz bandpass clicks as a function of the interclick interval (ICI) and the number of clicks in the train. The magnitude of the MLD grew as the number of clicks in the train was increased from 1 to 32. While the MLDs tended to be larger at longer ICIs, the effect was mediated by changes in detectability in the homophasic conditions. For click trains consisting of 4-32 clicks, the improvement in detectability in the antiphasic conditions with increases in the number of clicks appears to be the result of integration of acoustic power, as is the case for the homophasic conditions. The absence of MLDs for short trains of high-frequency transients remains quite puzzling, since large MLDs are found with single, low-frequency transients.     

Maximum cross-correlated kurtosis-based unsaturated stochastic resonance and its application to bearing fault diagnosis

Considering the random impulses of mechanical noise and the limitations involved while identifying mechanical fault impulse signals via traditional measurement indices of signal-to-noise ratio, which require the characteristic frequency to be known in advance, this study proposes an adaptive unsaturated stochastic resonance method employing maximum cross-correlated kurtosis as the signal detection index. The proposed method combines the features of a cross-correlated coefficient to indicate periodic fault transients and those of spectrum kurtosis to locate these transients in the frequency domain. Actual vibration signals collected from motor and gear bearings subjected to heavy noise are used to demonstrate the effectiveness of the proposed method. Through a coarse tree-based machine learning method, the proposed method is verified to be more suitable for explaining the periodic impulse components of bearing signals, as compared to the ensemble empirical mode decomposition denoising method and unsaturated stochastic resonance using the kurtosis-intercorrelation index.     

A framework for the automated real-time detection of short tonal sounds from ocean observatories

Acquisition of acoustic data from ocean observatories is expected to play a key role for the long-term monitoring of marine mammals and anthropogenic noise. It typically requires processing of a large volume of acoustic data and it must rely on automated identification of signals. We present an algorithmic framework for the detection of short tonal sounds (e.g. cetacean calls, anthropogenic pings) intended to act as a first stage in a system for the automated real-time detection, classification, and localisation of acoustic sources. The algorithm was validated under a diversity of scenarios expected at ocean observatories. Using simulated signals that emulate a variety of cetacean call-types, perfect identification of signal position was obtained for signal to noise ratios of ?15 to ?5 dB, depending on the signal-type. Separation of real-world data segments with short tonal sounds (mainly cetacean calls) from segments with other sounds or noise resulted in Area Under the ROC Curve values between 0.96 and 0.98. The algorithm can be used to automatically identify cetacean calls and anthropogenic short tonal sounds much faster than in real-time, thereby reducing the burden put on data transmission, storage, or processing by classification and localisation algorithms.     

Male sperm whale acoustic behavior observed from multipaths at a single hydrophone

Sperm whales generate transient sounds (clicks) when foraging. These clicks have been described as echolocation sounds, a result of having measured the source level and the directionality of these signals and having extrapolated results from biosonar tests made on some small odontocetes. The authors propose a passive acoustic technique requiring only one hydrophone to investigate the acoustic behavior of free-ranging sperm whales. They estimate whale pitch angles from the multipath distribution of click energy. They emphasize the close bond between the sperm whale's physical and acoustic activity, leading to the hypothesis that sperm whales might, like some small odontocetes, control click level and rhythm. An echolocation model estimating the range of the sperm whale's targets from the interclick interval is computed and tested during different stages of the whale's dive. Such a hypothesis on the echolocation process would indicate that sperm whales echolocate their prey layer when initiating their dives and follow a methodic technique when foraging.     

Low frequency narrow-band calls in bottlenose dolphins (Tursiops truncatus): signal properties, function, and conservation implications

Dolphins routinely use sound for social purposes, foraging and navigating. These sounds are most commonly classified as whistles (tonal, frequency modulated, typical frequencies 5-10 kHz) or clicks (impulsed and mostly ultrasonic). However, some low frequency sounds have been documented in several species of dolphins. Low frequency sounds produced by bottlenose dolphins (Tursiops truncatus) were recorded in three locations along the Gulf of Mexico. Sounds were characterized as being tonal with low peak frequencies (mean?=?990 Hz), short duration (mean?=?0.069 s), highly harmonic, and being produced in trains. Sound duration, peak frequency and number of sounds in trains were not significantly different between Mississippi and the two West Florida sites, however, the time interval between sounds within trains in West Florida was significantly shorter than in Mississippi (t?=?-3.001, p?=?0.011). The sounds were significantly correlated with groups engaging in social activity (F=8.323, p=0.005). The peak frequencies of these sounds were below what is normally thought of as the range of good hearing in bottlenose dolphins, and are likely subject to masking by boat noise.     

Dual mechanisms in the perceptual processing of click train temporal regularity

Two experiments measured human sensitivity to temporal jitter in 25-click trains with inter-click intervals (ICIs) between 5 and 100 ms. In a naturalistic experiment using wideband clicks, jitter thresholds were a nonmonotonic function of ICI, peaking for ICIs near 40-60 ms. In a subsequent experiment, clicks were high-passed and presented against a low-frequency noise masker. Jitter threshold vs ICI functions lost the positive slope over short ICIs but retained the negative slope at long ICIs. The same behavior was seen in click rate discrimination tasks. Different processes mediate regularity analysis for click trains with ICIs above and below 40-60 ms.