Automated identification and clustering of subunits within delphinid vocalizations

作者：Frasier Kaitlin E^*; Henderson E Elizabeth; Bassett Hannah R; Roch Marie A

来源：Marine Mammal Science, 2016, 32(3): 911-930.

DOI：10.1111/mms.12303

• acoustic
• classification
• communication
• delphinid
• dynamic time warping
• network analysis
• unsupervised learning
• vocalization

Tonal vocalizations or whistles produced by many species of delphinids range from simple tones to complex frequency contours. Whistle structure varies in duration, frequency, and composition between delphinid species, as well as between populations and individuals. Categorization of whistles may be improved by decomposition of complex calls into simpler subunits, much like the use of phonemes in classification of human speech. We identify a potential whistle decomposition scheme and normalization process to facilitate comparison of whistle subunits derived from tonal vocalizations of bottlenose dolphins (Tursiops truncatus), spinner dolphins (Stenella longirostris), and short-beaked common dolphins (Delphinus delphis). Network analysis is then used to compare subunits within the vocal corpus of each species. By processing whistles through a series of steps including segmentation, normalization, and dynamic time warping, we are able to automatically cluster selected subunits by shape, regardless of differences in absolute frequency or moderate differences in duration. Using the clustered subunits, we demonstrate a preliminary species classification scheme based on rates of subunit occurrence in vocal repertoires. This provides a potential mechanism for comparing the structure of complex vocalizations within and between species.

• 出版日期2016-7