Learned bird songs often have a hierarchical organization. In the case of zebra finches, each bird’s song is made up of a string of notes delivered in a stereotyped sequence to form a “motif”, and motifs are repeated to form a song bout. During song learning, young males copy “chunks” of two or more consecutive notes from their tutors’ songs. These chunks are represented as distinct units within memory (during learning) and within motor systems (during song production). During song performance, motifs may deviate from the learned sequence by stopping short, starting late, or by skipping, inserting or repeating notes. We measured acoustic and temporal variables related to the respiratory and vocal physiology of song production and asked how they related to deviations from each bird’s “canonical” sequence. The best predictor of deviations from that sequence was the duration of the silent interval between notes, when inspiration normally occurs. Deviations from the canonical motif occurred less often after higher-pitched notes, perhaps because a high-low sequence forms a prosodic unit. Premature stops often followed louder and longer notes, suggesting that respiratory and muscular physiology influence the location of such stops. Boundaries between the learned chunks of a male’s motif predicted where and how often non-canonical starts occurred. Physiological and cognitive elements also interacted to define the segmentation of zebra finch song sequences. Long silent intervals between notes were associated both with physiology (inspirations) and with the cognitive boundaries of learned chunks – and hence with deviations from the canonical motif.

Subjects:  The subjects were 50 adult male zebra finches (Taeniopygia guttata) bred and raised in the Williams College colony. Birds were hatched in aviaries that housed at least three breeding pairs, and so could copy songs from their fathers or from other adult males. Young males in their own and other broods could also influence song development (Tchernichovski & Nottebohm, 1998) or be the source of song material, and both young and older females were present and could influence song development (Carouso-Peck & Goldstein, 2019). When birds reached maturity (approximately 90 days) they were removed from their natal aviaries and housed in single-sex cages, grouped by age cohort, in a colony room that also housed females and older males. All procedures were approved by the Williams College Institutional Animal Care and Use Committee (Williams College IACUC protocol WH-A).

Song Recording:   Recordings of female-directed song were obtained by placing a caged adult male (> 120 days) in a closed Lucite chamber (75 x 60 x 50 cm) lined on two sides, the top, and the bottom with acoustic foam. A caged adult female was placed outside one of the transparent chamber walls to elicit directed (courtship) song. A Marantz EC-7 dynamic microphone in the chamber was used to record songs onto cassette tape (using a Marantz PMD-201 cassette recorder) or digitally (Marantz PMD-670 digital recorder; 44 kHz, 16 bit). Songs recorded on cassette tape were then digitized using SoundEdit Pro (Macromedia), which was also used to produce sound spectrograms for visual analysis of song sequences. For greater temporal resolution, 128 point FFTs were used. Multiple song bouts and motifs from each male were obtained from at least two recording sessions on different days. The mean number of motifs recorded and scored for each male was 119 (range = 59-188). The large number of motifs was required because deviations from the stereotyped order, although they occurred in 72% of motifs scored, were relatively infrequent at each transition between notes (mean = 7.4%).

Scoring Motif Sequences:  Multiple bouts and motifs sung by each bird were examined and annotated in order to find that male’s “canonical motif”, defined as the longest stereotyped sequence of notes within his songs. Figure 1a shows sound spectrograms of a song bout that includes two canonical motifs (the sequence of 12 notes labeled A through N) as well as five truncated or altered motifs (notes O-G, OE, AD-N, A-K, H-K). The sound file from which this bout is drawn is included in the Supplemental Materials. Calls that occurred in other contexts and were only rarely appended to the song were not considered to be part of in the canonical motif. Zebra finch song units separated by silent intervals have traditionally been called “syllables”, and these syllables are often the units used in analyzing motifs. However, some syllables consist of two or more “notes”, which we define as sounds that are bounded by abrupt transitions in frequency or frequency modulation. Notes  are not necessarily separated by silent intervals. We used notes rather than syllables as the unit of sound analysis because deviations from the canonical motif sometimes occurred between notes rather than between syllables, and so. Each bird’s canonical motif was represented as a string of letters corresponding to the order of its constituent notes (see Figure 1a). The first note in each bird’s motif was designated with the letter A, the second note with the letter B, and so on; hence the letters designate the order within a bird’s song rather than the note type. When introductory notes or calls were appended immediately after the last song note in a motif, but were not then immediately (with 200 ms) followed by another motif, they were not included in the string of labeled notes and transitions or in the analysis. Using the resulting annotated strings, each transition between notes within each bird’s song (where a deviation from the canonical sequence might fall) was designated by the notes that flanked a silent interval (e.g., start -> A, A -> B, … , M -> N). We then tabulated deviations from the canonical motif (see Figure 1a for six examples of such deviations). The most common types of deviations were a) stops before the canonical sequence was completed (1722 instances at 199 different locations within 50 birds’ motifs) and b) starts at points other than the first syllable of the canonical motif (981 instances at 50 locations within 28 birds’ motifs). Less common deviations included i) repeated notes or short sequences of notes, ii) skipped notes, and iii) insertions of notes into the canonical sequence. As each of these three types of deviations was relatively rare, we combined them into a single category for analysis: c) “other” deviations (675 instances at 84 locations within 42 birds’ motifs). Different types of deviations sometimes occurred at the same location within a single bird’s motif (80 locations in 44 birds’ motifs). Each individual deviation (and each type of deviation) from the canonical motif occurred at a specific transition between notes within a bird’s motif. We calculated the number of times each type of deviation occurred at each transition point as well as the number of opportunities within the bird’s song for such a deviation to occur (the number of motifs that included at least one of the flanking notes). We could then determine the frequency of each type of deviation at every possible transition between syllables within a bird’s canonical motif. Figure 1b shows sound spectrograms of each of four birds’ canonical motifs. In addition, Figure 1b shows the frequencies of the three types of non-canonical transition as both 1) bars within the spectrogram and 2) arrows within a network diagram of notes and transitions. A subset of 904 motifs (including 7,638 notes) sung by 22 of the subjects were scored by two individuals to check the reliability of our scoring method. A total of 2,010 transitions within these motifs were scored as stops, starts, skips, repeats, or insertions (most of the stops and starts fell at the beginning or end of canonical motifs and were not deviations). We found 27 discrepancies in the scoring of these 2,010 transitions, which represents 98.7% agreement in how different individuals scored the notes and transitions. 

Learned Chunks and Breaks Between Them:  To investigate the role of cognitive or perceptual factors in defining deviation points, we needed to define the boundaries of the learned “chunks” of notes that were copied as a unit during song development. To do so, we used a subset of 21 males. These birds sang songs with related canonical motifs, sharing parts of a “core sequence”  but differing in some details of that sequence. By comparing sound spectrograms of these birds’ canonical motifs, we were able match notes and sequences (see the four related songs in Figure 1c). We determined the “core motif sequence” to be nine consecutive notes in the song of the male who sang the maximum number of notes common to the 21 songs (male DB 10; see Figure 2). To find copied chunks of notes, we compared the canonical motifs of the other 20 males’ songs to the core motif sequence and noted points where the order of the core sequence was not followed, because of omissions, additions, or changes in the order of notes. The points within each male’s canonical motif where breaks in core sequence began and ended were defined as “chunk boundaries” (depicted by downward arrows in Figure 2). Some males’ canonical motifs included notes prior to before the beginning of the core motif, and one chunk boundary occurred prior to the first note of the core motif (W39 in Figure 2; a note was inserted immediately prior to the first note of the core sequence).  Similarly, some songs did not include the final note(s) of the core motif, and so ended at a chunk boundary (e.g. Pk42 in Figure 2) . Notes from the core motif sequence were omitted in some songs (resulting in a chunk boundary where the omission occurred), and notes were inserted into the core motif in other songs (resulting in two chunk boundaries, one on either side of the insertion). Based on the analysis depicted in Figure 2, each of the 158 intervals within the portions of the 21 birds’ motifs that matched some portion of the core sequence were scored as having a chunk boundary either present or absent. Notes and transitions outside the core sequence were not included in analyses that considered the position of chunk boundaries.

Measuring the Characteristics of Transition Points:  We measured the acoustic characteristics of each transition point within each bird’s canonical motif (see Table 1 for the measurements and units). Because FFT-derived sound spectrograms have an inherent temporal uncertainty due to the windowing of data points, timing and duration measures were taken from the sound waveform in SoundEdit Pro. To determine the position of an interval within a bird’s canonical motif, we measured the time from the beginning of that canonical motif to the beginning of the interval, and then divided that number by the duration of the entire canonical motif. Each interval position measurement thus falls between 0 (the start of the motif) and 1 (the end of the motif). To determine the relative amplitude of each note, we first normalized the amplitude of a recording of each bird’s canonical motif, and then measured the mean amplitude of that entire motif and of each note within the motif using SoundAnalysisPro (Tchernichovski, Nottebohm, Ho, Pesaran, & Mitra, 2000); soundanalysispro.com). The average amplitude of a note was then divided by the average amplitude of the motif to obtain that note’s relative amplitude, which ranged from 0.7 to 1.35; values greater than 1 correspond to notes with higher-than-average amplitude, and values less than 1 correspond to notes of lower-than-average amplitude. We also used SoundAnalysisPro to measure the mean pitch of each note. We also classified the notes flanking each interval within a bird’s song into three categories (see Figure 3): “downsweep” (a harmonic note that decreased by at least 300 Hz in pitch from beginning to end, n = 307), “stack” (an unmodulated note with harmonics, n = 151), “high” (notes with a mean pitch higher than 2 kHz and no energy apparent below 1.75 kHz in a sound spectrogram, n = 57). Some zebra finch note classifications split stacks and downsweeps into two categories based on duration, and may also define compound syllables that consist of two note types, such as “stack-downsweep” (Sturdy, Phillmore, & Weisman, 1999). We scored such compound syllables as two different notes. If two successive notes formed parts of a single continuous syllable, the duration of the inter-note interval was 0 ms. We did not include the position of a motif within a song bout in our analysis of deviations from individual birds’ canonical sequence. We observed deviations from canonical motif sequences in all possible positions within bouts: the first, last, and internal motifs, within bouts that included multiple motifs, in single-motif bouts, and in motifs preceded by varying numbers of introductory notes. However, we cannot rule out the possibility that a study of how bout structure is related to deviations might find that some aspects of bout structure (Hyland Bruno & Tchernichovski, 2019), such as position within a bout or number of introductory notes, could be correlated with the occurrence of deviations from a bird’s canonical motif sequence.

R, with the following packages: ggplot2, Rcpp, Matrix, lme4, reshape2, sciplot, lattice, plyr, magrittr, MuMIn, Amelia, pscl, readr, performance, dplyr, tidyverse, ggeasy, car, glmmTMB, bbmle, GLMMadaptive, sandwich, lmtest, MASS, DescTools (all are open-source)