Abstract:

The performance of two well-known recognition techniques in speech adapted to the automated recognition of bird song units from continuous recordings is studied. The advantages and limitations of dynamic time warping (DTW) and hidden Markov models (HMMs) are evaluated on a large database of songs of two species: zebra finches (Taeniopygia guttata) and indigo buntings (Passerina cyanea), which have different types of vocalizations and have been recorded under different laboratory conditions. The recognition performance of these methods is also assessed with regard to song signal representation (including representations commonly used in speech recognition), model structure, and sensitivity. Depending on the quality of recordings and song complexity, the DTW-based technique gives from excellent to satisfactory results. Under more challenging conditions, such as noisy recordings or the presence in bird vocalizations of confusing transient calls, HMMs can significantly outperform DTW but requires more training examples. Even though the HMMs perform usually quite well, they often misclassify short transient calls and song units with more variable structure. To address these and other weaknesses of the studied techniques, new approaches to analyze transient and variable bird vocalizations are discussed. [Work supported in part by U.S. Army Research Office and NIH.]