Using self-organizing maps to classify humpback whale song units and quantify their similarity

Abstract

Classification of vocal signals can be undertaken using a wide variety of qualitative and quantitative techniques. Using east Australian humpback whale song from 2002 to 2014, a subset of vocal signals was acoustically measured and then classified using a Self-Organizing Map (SOM). The SOM created (1) an acoustic dictionary of units representing the song's repertoire, and (2) Cartesian distance measurements among all unit types (SOM nodes). Utilizing the SOM dictionary as a guide, additional song recordings from east Australia were rapidly (manually) transcribed. To assess the similarity in song sequences, the Cartesian distance output from the SOM was applied in Levenshtein distance similarity analyses as a weighting factor to better incorporate unit similarity in the calculation (previously a qualitative process). SOMs provide a more robust and repeatable means of categorizing acoustic signals along with a clear quantitative measurement of sound type similarity based on acoustic features. This method can be utilized for a wide variety of acoustic databases especially those containing very large datasets and can be applied across the vocalization research community to help address concerns surrounding inconsistency in manual classification.

Cultural revolutions reduce complexity in the songs of humpback whales

“…into a numeric sequence of units using a repertoire 'dictionary' as a guide [20]. This dictionary was created based on acoustic features of a representative subset of units from the eastern Australian population [20]. Song cycles from each year were transcribed and analysed [20], with a 'song cycle' defined here as a stereotyped set of themes sung in a specific order [13,15,16].…”

“…This dictionary was created based on acoustic features of a representative subset of units from the eastern Australian population [20]. Song cycles from each year were transcribed and analysed [20], with a 'song cycle' defined here as a stereotyped set of themes sung in a specific order [13,15,16]. Thirty-six song cycles were transcribed from each year, except for 2006 (12 cycles) and 2007 (four cycles) owing to lack of high-quality recordings.…”

“…This analysis was weighted (b ¼ 1) to account for acoustic similarity of units [20,26]. To ensure that qualitative theme assignments were robust, LSI was calculated between every possible pair of phrase repetitions in a given year and hierarchically clustered [20]. See Garland et al [26] for details on the LSI analysis, acoustic dictionary and unit classification.…”

Bidirectional Interactions With Humpback Whale Singer Using Concrete Sound Elements

“…The automatic approaches clearly help to better describe these songs. A very large panel of methods were proposed, for example, based on the information entropy (Suzuki et al, 2005), on a threshold on the Fourier coefficients (Mellinger, 2005), on the extraction of the edge contour (Gillespie, 2004), on the analysis of the mel-frequency cepstral coefficient (MFCC) (Pace et al, 2010) or the wavelet coefficients (Kaplun et al, 2020), or with the use of artificial neural networks (Allen et al, 2017;Mohebbi-Kalkhoran, 2019). To go further with the perception of these vocalizations, new representations were suggested and were very interesting to better extract the similarities of units in these songs, especially based on colored pictograms (Rothenberg and Deal, 2015).…”

Use of recurrence plots for identification and extraction of patterns in humpback whale song recordings

“…where the unit is placed, so that other hierarchical levels of organization can help define how much variation is allowed in a single unit type. Note that even when studies carry on unit annotation using computational methods, the results are usually validated using manual classification (Pace et al (2010), Garland et al (2013), Allen et al (2017)).…”