Speech recognition meets bird song: A comparison of statistics‐based and template‐based techniques Free

Pattern recognition technology that has been developed for recognizing units of human speech can often be adapted for both recognition and analysis of animal vocalizations. This paper discusses two types of speech recognition algorithms, template based and statistics based, with respect to their ease of deployment and potential application to the objective, quantitative analysis of animal vocalizations. Implementations of the two types of algorithms have been compared using a large database of song units recorded from two song bird species. The algorithms exhibit different strengths and weaknesses. The template‐based dynamic time‐warping algorithm provides quantitative sound comparisons that are directly useful to a researcher, but selection of training materials depends on expert knowledge. The statistics‐based hidden Markov model algorithm requires more training data, but usually performs better in noisy environments and with more variable vocalizations. While both algorithms are accurate in restricted domains, recognition performance could be improved if it were based on species‐specific features extracted from the acoustic input. [Work supported by NIH 1‐F32‐MH10525 and ARO DACA88‐95‐C‐0016.]