Figure - uploaded by Michel Kreutzer
Content may be subject to copyright.
Source publication
A crucial step in the understanding of vocal behavior of birds is to be able to classify calls in the repertoire into meaningful types. Methods developed to this aim are limited either because of human subjectivity or because of methodological issues. The present study investigated whether a feature generation system could categorize vocalizations...
Contexts in source publication
Context 1
... and Roy (2009) introduced a method to automati- cally create efficient acoustic features for specific audio-clas- sification problems. This approach has been shown successful when applied to dog vocalizations (Molnár et al., 2008). Therefore, the approach by Pachet and Roy (2009) was used to generate features for the current five call-type problem. ...Similar publications
Localization and tracking of marine animals can reveal key insights into their behaviors underwater that would otherwise remain unexplored. A promising nonintrusive approach to obtaining location information of marine animals is to process their bioacoustic signals, which are passively recorded using multiple hydrophones. In this paper, a data proc...
Citations
... To identify the different themes, a trained researcher visually analyzes the spectrogram of the song and identifies some patterns of the sound [7]. In the chingolo, typical features that must be analyzed are the shape, the type of modulation, and the number of notes. ...
... A visual comparison between the themes of 1966, and the ones we measured in 2020, allowed us to identify that there are similar themes. However, similarity clustering using visual analysis is subjective, and depends on the researcher's experience identifying patterns as well as it is conditioned by his/her biases [7,8]. ...
In this work, we use tools from nonlinear dynamics to generate synthetic bird songs with frequency modulations compatible with sketches reported in a study of chingolo (Zonotrichia capensis) songs in 1966. Using machine learning tools, we conclude that some of the sketches correspond to themes that are still sang in the same region, five decades later.
... Life recordings: This category of datasets was generated by the researchers specifically for their research. It is made up of recordings of the subject of interest either in their natural habitat (Allen et al., 2017;Briggs et al., 2012;Ibrahim et al., 2019;LeBien & Ioup, 2018;Roch et al., 2011;Shamir et al., 2014), or in a controlled environment such as recording rooms and laboratories (Giret et al., 2011;Oikarinen et al., 2019;. In some cases, a recording device was attached to the animals while for humans a Bluetooth stethoscope was used to obtain recordings of heart sounds. ...
This study is a systematic review of literature on the classification of sounds in three domains - Bioacoustics, Biomedical acoustics, and Ecoacoustics. Specifically, 68 conferences and journal articles published between 2010 and 2019 were reviewed. The findings indicated that Support Vector Machines, Convolutional Neural Networks, Artificial Neural Networks, and statistical models were predominantly used in sound classification across the three domains. Also, the majority of studies that investigated medical acoustics focused on respiratory sounds analysis. Thus, it is suggested that studies in Biomedical acoustics should pay attention to the classification of other internal body organs to enhance diagnosis of a variety of medical conditions. With regard to Ecoacoustics, studies on extreme events such as tornadoes and earthquakes for early detection and warning systems were lacking. The review also revealed that marine and animal sound classification was dominant in Bioacoustics studies.
... This method's main advantage is that it avoids categorizer's subjectivity, and it can be highly accurate (e.g., Pérez-Espinosa et al. 2015;Tchernichovski et al. 2000). On the other hand, it is limited by the range of feature sets used, which need not be applicable to a particular categorization (Giret et al. 2011). Parameters are extracted either manually or automatically using various commercial sound programs such as Avisoft-SASLab Pro (Raimund Specht, Berlin) or Raven Pro (Center for Conservation Bioacoustics 2011) custom programs such as the LMA, which focus on atonal sounds (developed by K. Hammerschmidt; see Schrader and Hammerschmidt 1997), or Sound-analysis R packages, such as Seewave (Sueur et al. 2008), EchoviewR (Harrison et al. 2015), or warbleR (Araya-Salas and Smith-Vidaurre 2017). ...
... Aside from these commonly used methods, Giret et al. (2011) proposed an approach known as feature generation method. This classification method combines human categorizers' involvement with automatic feature measurement. ...
... This classification method combines human categorizers' involvement with automatic feature measurement. In particular, visual categorization is used as a training database subsequently used to train and test a classifier by a supervised classification technique (Giret et al. 2011). Feature generation outperforms methods based on standard features in studies on animal vocalization (Giret et al. 2011;Larranaga et al. 2015;Molnár et al. 2008). ...
Parrots belong to the handful of animal species capable of vocal production learning. They are considered to be open-ended learners with complex and variable vocalizations. It is not known, however, to what extent their repertoires are shared among individuals within a group or between vocally isolated individuals or groups. In study 1, we mapped the repertoire of four captive African gray parrots (Psittacus erithacus) using a combination of three acoustic analyses. In study 2, we compared the repertoire of two female African gray parrots from two different laboratories who had never been in vocal contact with each other or any member of the other parrot’s social group. Results of study 1 showed a relatively large agreement between all three methods used to analyze the vocalizations. Almost three quarters (72.8%) of categories determined by visual-acoustic analysis were confirmed by at least one of the two computer-aided methods used, i.e., by spectrographic cross-correlation and/or a multiparametric statistical method. In study 2, we found a relatively large proportion of calls shared between the repertoires of the two gray parrot subjects. Over half and over a fourth of calls produced by parrots with the smaller and the larger repertoire, respectively, were shared between the two. No previous study identified such a large proportion of intergroup shared calls within this parrot species. It seems that some calls tend to reappear in vocally isolated groups based on inherited predispositions, similarly as has been documented in songbirds.
... Thus, our gaze following data might represent false negative results and should be considered with precaution. Overall, the previous positive findings on the ability of parrots to imitate a conspecific or human experimenter (Heyes and Saggerson, 2002;Moore, 1992;Mottley and Heyes, 2003), to respond to humans' intentional actions (Péron et al., 2010), and to use proximal human pointing gesture (Giret et al., 2011), may cast some doubt on whether the poor responsiveness or chance level performance of our parrots in the social domain reliably reflects their actual cognitive capacity. ...
The study of comparative cognition aims to elucidate the evolution of different domains of cognition. By comparing and contrasting closely and distantly related species we can identify the selective pressures that may have lead to their specific cognitive traits. Causal cognition is considered a more complex cognitive trait and is defined as how individuals understand cause and effect relationships. Causal relationships can be learned associatively, but there also appears to be some species that have an intuitive understanding of the mechanisms that underlie causal relationships. Parrots represent a relatively understudied and diverse group of birds that display complex cognitive traits. Studying causal cognition in parrots therefore provides information on a broad range of species that are adapted to a varied range of niches, which vastly broadens our perspective of cognitive evolution. I studied whether different species of parrots were capable of using understanding of causal relationships to guide their behaviour, and specifically whether this causal understanding was driven by learned associations or by complex cognitive processes based on underlying mechanisms.
I tested four species of parrots from the Psittacidae family: three species from the Arinae subfamily, Ara ambiguus (great green macaws), Ara glaucogularis (blue throated macaws) and Primolius couloni (blue headed macaws), and one species from the Psittacinae subfamily, Psittacus erithacus (African grey parrot). In the first chapter I tested all four species on an established test battery to evaluate a broad range of cognitive factors from both the physical and social domain. I was able to compare the parrots’ results to both apes and monkeys that had previously completed the same test battery. The parrots did poorly, but it was likely due to issues with task design that made the tests especially difficult for the parrots compared to primates. Following this, in the second chapter, I tested the two Ara species on an established causal cognition test: the trap-tube. Although many subjects were relatively successful in the task, the macaws showed a preference to use learned rules to find a solution rather than an understanding based on the underlying mechanism of the apparatus. Finally, in the third chapter, I once again tested the two Ara species in another causal cognition test: the stone- dropping task. There was a division between the two Ara species in their approach to solve the task. It appears that most of the Ara ambiguus were able to solve the task through exploratory behaviour alone whereas the Ara glaucogularis appeared to require information on the underlying mechanism of the apparatus to solve it.
The first two chapters entailed choice tasks, where subjects were required to pick between two or three options to find a reward. I discuss how this type of testing is not suitable for exploring aspects of complex cognition as most subjects can solve these tasks using simple learned rules. However, based on results of the second chapter, I discuss how if choice tests must be used then there are ways to improve subjects’ motivation to not use simple learned rules. Based on the results of the third chapter, I was able to discuss in more detail what kind of causal understanding the two Ara species used. I came to the conclusion that although they may have the capacity to use causal understanding based on the underlying mechanisms of apparatuses, they do not appear to have a drive to do so, favouring learned causal associations where possible. Finally, I contextualise these results on what we know about these parrots ecology and suggest that a causal understanding based on functional mechanisms may not be a factor that is especially vital for macaw cognition.
... Attribution (which animal vocalized). Automated classification has been done in various species, including rodents (Kobayasi & Riquimaroux, 2012;Soltis, Alligood, Blowers, & Savage, 2012), frogs (Pettitt, Bourne, & Bee, 2012), birds (Giret et al., 2011), bats (Prat, Taub, & Yovel, 2016) and primates (Fuller, 2014;Hedwig, Hammerschmidt, Mundry, Robbins, & Boesch, 2014), including marmosets (Agamaite, Chang, Osmanski, & Wang, 2015;Turesson, Ribeiro, Pereira, Papa, & De Albuquerque, 2016;Zhang, Huang, Gong, Ling, & Hu, 2018). In previous marmoset studies, it was possible to detect the vocalizations by amplitude thresholding of the band-or high-passed audio signal. ...
We introduce an end-to-end feedforward convolutional neural network that is able to reliably classify the source and type of animal calls in a noisy environment using two streams of audio data after being trained on a dataset of modest size and imperfect labels. The data consists of audio recordings from captive marmoset monkeys housed in pairs, with several other cages nearby. Our network can classify both the call type and which animal made it with a single pass through a single network using raw spectrogram images as input. The network vastly increases data analysis capacity for researchers interested in studying marmoset vocalizations, and allows data collection in the home cage, in group housed animals.
... Studies focusing on this level of variation are often assumptive or subjective in methodology (Milinski 1997). Non-automated visual classification of bioacoustics signal is prone to perceptual bias by the observer (Giret et al. 2011;Cholewiak et al. 2013), so computational methods can help by taking into account uncertainty in the estimates of similarity and the classification of bioacoustic signals (Ranjard et al. 2008(Ranjard et al. , 2015. ...
Geographic patterns of song variation are common in passerines and can develop as a consequence of the mechanisms of song acquisition and dispersal. In particular, the timing of dispersal relative to the sensory learning phase and the time of song crystal-lization is important. For example, when the sensory phase continues after dispersal or when males learn new songs every breeding season, i.e. open-ended learner, neighborhoods can develop where males share their songs. In this study, we utilize a unique, comprehensive dispersal and song recording dataset to investigate the existence and development of micro-geographic song variation in a wild passerine population. Machine learning song analysis methods allow us to overcome perceptual bias in the classification of the songs of New Zealand hihi (Notiomystis cincta). We show that males share more song elements of their repertoire with their neighbors than with more distant males or with males from the same natal area, implying that repertoire is acquired post-dispersal. Finally, we suggest that high levels of male competition have driven the development of post-dispersal vocal learning behavior in this species.
... Encoding songs as sequences of syllables can be done either by ear and spectrogram visual analysis or by specific automatic algorithms (Ranjard and Ross 2008). Methods of automatic syllable identification directly from the recordings are designed and used for recordings from birds in captivity (Giret et al. 2011) and good field recordings of species with comparatively simple and small repertoires and short songs such as hihi (Ranjard et al. 2015) and whitecrowned sparrow (Ranjard and Ross 2008). Undoubtedly, these methods are extremely useful in analysis of large amount data and big samples of individual recordings. ...
Among songbirds there are species with rich and variable repertoires and very complex sequences of sounds that are hard to analyse. Here, we investigate repertoire and syntax organization in the large-billed reed warbler advertising vocalization. We provide the first detailed qualitative and quantitative description of a song of a recently rediscovered species with the analysis of song syntax by clustering. The song of A. orinus combines discrete and continuous types of singing. Although the vocalization can be produced continuously for minutes without any significant pauses, it is easy to split it into distinctive songs using the criterion of 500 ms pause. Songs themselves are grouped into song classes. Both songs and song classes are unique to each individual, though several simple notes could be shared between males. Linear syntax is found both in sequences of syllables and in sequences of songs. Each song starts with initiator syllable types. We also discovered the presence of the transitional songs that sometimes separate songs from different classes in the vocal stream. The repertoire size in our sample varies from 69 to 147 syllable types and from 10 to 24 song classes, and does not depend on the duration of recording. Only one male’s repertoire was estimated entirely. This male was singing alone while other males sang in the presence of other individuals. We discuss how the presence of listeners may affect the repertoire and temporal and sequential organization of the song.
... Several recent studies have used quantitative methods to examine the vocal repertoire of a diverse array of species, including primates (Hedwig et al., 2014;Fuller, 2014), rodents (Kobayasi and Riquimaroux, 2012;Soltis et al., 2012), birds (Giret et al., 2011), and frogs (Pettit et al., 2012). These kinds of analyses are critical to understanding the types of vocalizations produced by a particular species and for carrying out behavioral studies that investigate properties such as perceptual boundaries between vocalization types as well as neurophysiological studies that investigate neural coding mechanisms for vocalizations. ...
The common marmoset (Callithrix jacchus), a highly vocal New World primate species, has emerged in recent years as a promising animal model for studying brain mechanisms underlying perception, vocal production, and cognition. The present study provides a quantitative acoustic analysis of a large number of vocalizations produced by marmosets in a social environment within a captive colony. Previous classifications of the marmoset vocal repertoire were mostly based on qualitative observations. In the present study a variety of vocalizations from individually identified marmosets were sampled and multiple acoustic features of each type of vocalization were measured. Results show that marmosets have a complex vocal repertoire in captivity that consists of multiple vocalization types, including both simple calls and compound calls composed of sequences of simple calls. A detailed quantification of the vocal repertoire of the marmoset can serve as a solid basis for studying the behavioral significance of their vocalizations and is essential for carrying out studies that investigate such properties as perceptual boundaries between call types and among individual callers as well as neural coding mechanisms for vocalizations. It can also serve as the basis for evaluating abnormal vocal behaviors resulting from diseases or genetic manipulations.
... Because it was not possible to tune all reasonable combinations of the six parameters and eight feature sets, we did some preliminary exploration with the 100 training set recordings to establish a set of default parameter values (those marked with an asterisk in Table 2). We then tuned the parameters in a greedy manner (Goodrich and Tamassia, 2002) using the values and in the order listed in Table 2, which is the order that they are employed in the workflow. Ultimately, the default values obtained in our preliminary exploration remained optimum. ...
... Therefore, it is critical to reliably measure sound similarities especially when a large number of recordings are being analyzed. Moreover, one needs to minimize any bias that could be introduced by the specific sensory attributes of human experts, which could preclude reproducibility of the analysis (Milinski, 1997;Giret et al., 2011). For example, inconsistency across investigators has been reported in the classification of humpback whale songs (Cholewiak et al., 2013). ...
... We report relatively low agreement between human expert classifications on the test dataset. Therefore, there is a level of uncertainty in the grouping of the songs preventing the use of a single benchmark expert classification to (i) optimize the song variant classifier (Trawicki et al., 2005;Tao et al., 2008) and (ii) select the best acoustic parameters using, for example, a feature generation approach (Giret et al., 2011). We, therefore, define an unsupervised showing the pairwise song distances obtained from encoded spectrogram alignments of a subset of 469 songs. ...
Human expert analyses are commonly used in bioacoustic studies and can potentially limit the reproducibility of these results. In this paper, a machine learning method is presented to statistically classify avian vocalizations. Automated approaches were applied to isolate bird songs from long field recordings, assess song similarities, and classify songs into distinct variants. Because no positive controls were available to assess the true classification of variants, multiple replicates of automatic classification of song variants were analyzed to investigate clustering uncertainty. The automatic classifications were more similar to the expert classifications than expected by chance. Application of these methods demonstrated the presence of discrete song variants in an island population of the New Zealand hihi (Notiomystis cincta). The geographic patterns of song variation were then revealed by integrating over classification replicates. Because this automated approach considers variation in song variant classification, it reduces potential human bias and facilitates the reproducibility of the results.
