Hidden Markov models for the analysis of animal vocalizations.

Hidden Markov models (HMMs) are commonly used to model sequential data, including time series such as vocalizations. HMMs are uniquely suited to model time series because of their ability to implicitly align samples and to be linked together to create flexible recognition patterns for vocalizations with repeating patterns, such as bird and whale songs. HMMs require less human interaction and manual tuning for successful recognition and are inherently more noise resistant than template‐based techniques. Due to their statistical basis, HMM‐based recognition systems can incorporate additional statistical models. One such example is a language model, which can be designed to include a priori knowledge about the structure of the vocalizations, such as syllable repetition patterns. HMMs have become the most popular recognition model in speech processing, and this experience has been applied to a variety of vocalization analysis tasks including individual identification, song classification, call type recognitio...