It is generally difficult to decide if two performances of a behavior are of the same pattern and what are the natural units of a behavioral process. Assessment of song imitation raises both issues. First, comparing the song of a tutor-bird with the emerging song of its pupil: it is difficult to decide if (and at what stage) the two performances are similar enough to be considered as 'imitation'. Second, although it is relatively easy to divide a mature song into 'syllable units', during the imitation process the acoustic structure of syllables is vague and is constantly changing. Therefore even the first step of partitioning a developing song into distinct sounds is a difficult task. To address those subjects we developed a fully automated procedure that measures parametrically the similarity between songs. The procedure uses an analytical framework of modern spectral analysis to characterize the acoustic structure of a song.

This analysis provides a superior sound spectrogram that detects frequency traces even during early stages of vocal ontogeny, when sounds are vague and variable. The procedure then reduces the spectral image to a set of simple acoustic features. Based on these features, it detects similar sections between songs automatically as illustrated in Figure 1.

This technique allows us to analyze the process of song imitation as it occurs from moment to moment: We induced the rapid onset of song changes leading to song imitation in zebra finches and used our automated procedure for tracking trajectories of these changes from the initial untutored sounds and until a model match was achieved. We found that exposure to a song model induced marked generation of structured sounds during the second day of training. The temporal pattern of sounds then slowly transformed from repetitive to serial production of syllables.

Figure 1. Our automatic procedure examines the similarity between two sounds by constructing a similarity matrix that combines four acoustic features: pitch, Wiener entropy, FM and spectral continuity. The algorithm sets a statistical threshold to distinguish between non-similar (black) and similar (gray) sections of the similarity matrix. It then selects the most-similar sections (colored lines), and examines fine-grained assessment of the similarity across the acoustic features.

Tracking this transition revealed that different sounds of the song model were not copied separately, but often differentiated, through stereotyped trajectories, from successive renditions of the same prototype sound. We suggest that song imitation is a process of generation and differentiation of primitive prototype sounds. The natural units of the process are therefore the operations that the bird performs on sounds so as to match them to the model.

Analysis of imitation trajectories in terms of such operations allows us to text hypotheses regarding the underlying mechanism of vocal imitation: Importantly, there were several cases where imitation did not proceed by gradual reduction of acoustic differences between the sounds of model and pupil. We explain those deviations by synthesis of our findings with previous analysis of the non-linear dynamics of song production in the bird's vocal organ. Based on those findings, we present a framework for a new theory of song imitation.

© 2000 Noldus Information Technology b.v.