Figure 1.
Selection of tutor and pupil songs for similarity analysis.
A) Example of tutor song showing stereotyped motifs. B) Examples of pupil song bouts, showing variable song structure including motifs and other irregular vocal elements. Some segments of pupil song are more similar to the tutor motif than others. Green bars represent apparent motifs. C–F) The effect of manual (blue) versus automatic (black) selection of song segments on similarity analysis (computed with Sound Analysis Pro, SAP). C) The songs of each bird are compared to other songs of the same bird (self-similarity) or to songs of other adult birds in the colony (cross-similarity). Each point in (C) plots the self-similarity vs cross-similarity score for one bird based on the acoustic similarity of the songs. D) Contrast between self- and cross-similarity, computed for acoustic similarity scores. Each point is one bird. E) Sequence self-similarity versus sequence cross-similarity. F) Contrast between self- and cross-similarity, computed for sequence similarity scores. Each line in figures D,F connects results from the same birds, carried out either by manually or automatically selecting song segments. G) Examples of inconsistent segmentation of song into syllables and silent gaps. For panels A-B, G: top, song spectrogram; middle, segmentation of the song into syllables (red bars); bottom, sound amplitude (log power).
Figure 2.
Spectral features for representation of song.
We consider the following 6 features for representing song acoustic structure: Top to bottom: gravity center, spectral width, Weiner entropy, pitch goodness frequency modulation and pitch.
Figure 3.
Computation of acoustic and sequence similarity from the similarity matrix.
A) Pupil song segment (spectrogram at top), tutor motif (at left) and matrix of similarities between all time points in the two songs. For each of the tutor syllables at left, the red diagonal line represents the best match in the pupil song. The yellow circle marks the diagonal with the highest scores (computed as the integral along the diagonal). This represents the selected best match to that tutor syllable. B–E) The best-matched tutor syllable and section of pupil song are removed from the similarity matrix. B) The best matches to the remaining tutor syllables are recomputed. The yellow circle marks the diagonal with the largest score. C–E) The best-matched tutor syllable and pupil song section are removed from the similarity matrix and the process is reiterated until all tutor syllables have been matched. F) Computation of sequence score. Top panel: For syllable ‘d’ the best matching is shown by a red diagonal and, for illustration, the matched pupil song fragment is denoted here by dpupil. The algorithm then measures the similarity between the next tutor syllable (‘e’) and the fragment of pupil song which follows dpupil. This area of interest is marked by a dashed red box, below and to the right of the red diagonal. The algorithm finds the highest-scoring diagonal within the area of interest, denoted here by cyan diagonal, and this is the partial sequence score for syllable d. Other panels: partial sequence scores for the other syllables. The overall sequence score is the average over the partial sequence scores for all syllables.
Figure 4.
Comparison of different methods of measuring acoustic similarity.
A) Selecting the optimal set of features. For each bird we measured the similarity of extracted song bouts to its own song motif (self-similarity) and to the motif of other birds (cross-similarity). These were computed using different combinations of spectral features (indicated with different colors and symbols). Inset: The contrast between self-similarity and cross-similarity, shown for each different subset of features tested. B-E) The SI algorithm yields higher contrast than the SAP software. Acoustic (B) and sequence (D) self-similarity versus cross-similarity computed using the Similarity Index (SI) algorithm with the optimal features (red), using the SI algorithm with the set of features used by SAP (cyan), and using SAP software (blue). The contrast was significantly larger using the SI algorithm with optimal features, both for the acoustic similarity scores and sequence similarity scores (C and E, respectively).
Figure 5.
Changes in acoustic and sequence similarity through vocal learning.
Quantification of tutor imitation in a set of four juvenile birds at early, middle and late stages of vocal learning. For each juvenile bird, song similarity to the tutor song was quantified on days corresponding to 60, 75 and 90 days of age (days post hatch, dph). Acoustic similarity and sequence similarity were quantified separately, and developmental changes were computed by subtracting the similarity scores at 60 dph. (A) Change in acoustic similarity at 75 and 90 dph. Shown are scores computed using the SI algorithm (black) and the SAP algorithm (grey), both of which show significant developmental increase in similarity to tutor song. B) Change in sequence similarity at 75 and 90 dph. The SI algorithm reveals significant development of sequence imitation. The SAP sequence scores exhibit no significant correlation with age. C–F) For each of the four birds songs early and late in development. Each figure corresponds to one bird top to bottom: tutor song; bird song recorded at age 90 dph; two examples of bird song recorded at age 60 dph.