Limits on Reacquisition of Song in Adult Zebra Finches Exposed to White Noise

Subjects

Male zebra finches used in this experiment were bred and raised in group aviaries, each of which contained five adult breeding pairs, on a 14/10 light/dark schedule at 25°C. At 55 d post-hatch (juvenile stage), subjects were moved to small flight cages in which they were housed in groups of three. This allowed ample exposure to their father's song to ensure normal song development (Böhner, 1990). At 120 d post-hatch (adulthood), subjects were housed individually in these cages, which were placed within sound-attenuated boxes, and WN exposure began.

WN exposure

Twelve sound-attenuating boxes (23 × 20 × 26 inches) were built in-house and consisted of double plywood walls separated by acoustic foam. Two WN generators (Quan-Tech 420 and General Radio Company 1382; both from Tucker Davis Technologies, Garland, TX) were used to present WN, driven by two six-channel amplifiers (NAD 916). Each sound-attenuating box was equipped with a four-inch, 200 W two-way speaker (Houston Acoustic 4202), mounted on the top side of one edge. To test whether the noise exposure was consistent at all audible frequencies (250 Hz to 20 kHz), WN was recorded using a reel-to-reel tape recorder (TEAC X-300) and analyzed using a digital sound spectrograph (Kay Elemetrics DSP 5500). The WN was found to be flat within 5 dB at all frequencies, measured at 100 dB. Because of near-field acoustics, the precise amplitude of the WN was not constant at all points in the sound-attenuating boxes. Levels were thus set so that the minimum amplitude was 100 dB as measured at five different points in each box with a hand-held sound level meter (Realistic 33-2050). In practice, this meant that the amplitude 2-3 cm from the speaker ranged from 116 to 120 dB. The maximum volume at the level of the perch was ∼103 dB. This level of WN does not result in damage to hair cells in zebra finches (Ryals et al., 1999; Sarah Woolley, personal communication).

Four groups of birds were exposed to different schedules of WN. Groups 1 and 2 were both exposed to a mixture of “interrupted” and “continuous” WN schedules. Interrupted refers to a schedule on which song was recorded once weekly for 1 hr or until 30 song samples were recorded. During recording, WN was turned off, allowing the subject to hear its own vocalizations (notwithstanding any short-term effect of WN on hearing; see below). These weekly, hour-long recording sessions were the only time during which noise was not present on the interrupted schedule. For subjects on the continuous schedule, WN was simply never turned off, and vocal behavior of the birds was not recorded. Our initial motivation for including the interrupted condition was to allow us to observe changes in song over the course of exposure to WN. However, we observed no substantial change in song over extended periods during which subjects were exposed to interrupted noise, and these birds were therefore switched to a continuous schedule because the hourly interruption of noise was apparently preventing changes in song. Whether such brief exposure to song can actually serve to counteract the effects of WN exposure observed here remains an open question for future research. Groups 3 and 4 were run later and were maintained on continuous noise from the outset. In cases in which siblings from the same clutch were run, they were evenly distributed across treatment groups.

The conditions were: birds in group 1 were maintained on an interrupted schedule for 2 months, after which they were switched to a continuous schedule for 4 months (n = 6). Group 2 birds were maintained on an interrupted schedule for 6 months, after which they were switched to a continuous schedule for 4 months (n = 5). Birds in group 3 were maintained on a continuous schedule for 5 months (n = 6). Group 4 birds were maintained on a continuous schedule for 7 months (n = 5). Finally, a control group of five birds was recorded on two different occasions, 3 months apart without any treatment to establish a baseline for measures of temporal aspects of song (see below).

Song recording

Song was recorded using automated recording software (Avisoft SASLab). Microphones (Shure BG 1.1) were positioned inside the sound-attenuated boxes and each channel independently preamplified using a mixing board (Mackie 1604-VLZ). Song was digitized using an eight-channel soundcard (SEK'd ARC 88), allowing simultaneous monitoring of eight birds. Birds were recorded on two separate occasions (1 week apart) before the onset of WN exposure. After WN onset, weekly recording sessions (interrupted conditions only) lasted 1 hr, or until 30 samples of song were recorded. On some occasions, up to 5 hr would elapse without 30 samples being recorded for a given subject. Rather than allow a lengthy interruption of WN exposure, recording was discontinued after 5 hr and WN was restored. After WN was discontinued, recordings were made daily for 2 weeks, then weekly for at least 100 d. The time points discussed in Results include the pre-WN recordings (PRE), the first day out of WN (D1), and after 100 d out of WN (D100).

Tutoring

Potential tutors were recorded, and spectrograms made of their songs. These were compared with the subjects' songs, and pairings were selected to maximize dissimilarity between the tutors' songs and the subjects' PRE songs. In this way, any observed increases in similarity to a tutor song could be distinguished from recovery of PRE song more clearly. Birds in group 1 were tape-tutored using computer playback for 2 hr per day. Sequencing software (Cakewalk Pro Audio 8) was used to present two song samples every 10 min during each hour-long tutoring session, one in the morning and one in the afternoon. Birds in groups 2, 3, and 4 were presented with live tutors immediately after their initial recording on D1 after removal from continuous WN. Tutors were housed continuously with subjects, except during recording sessions.

Evoked potential recordings

To determine whether long-term exposure to WN had lasting effects on hearing, we recorded evoked potentials from auditory brainstem in 15 birds housed in continuous WN under the same conditions as experimental birds for 1 month. Recordings were made at 1 hr, 3 d, and 1 week after removal from WN (five birds in each condition) using procedures modeled after those of Woolley et al. (2001).

Auditory thresholds were determined by evoked potential recordings from auditory brainstem during the presentation of 10 msec pure tones (rise/fall time of 1 msec) at frequencies of 0.25, 0.5, 1.0, 1.5, 2.0, 3.0, 3.5, 4.0, 5.0, and 6.0 kHz, presented at a rate of 10 per second. Stimuli were generated using custom software (Brandon Warren, University of Washington, Seattle, WA) and delivered with a free field speaker (RCA PROX55AV), placed at a 90° angle to the head at a distance of ∼10 cm. Evoked responses were digitized using the Line In input to the soundcard and recorded by the same software used for stimulus generation.

Subjects were anesthetized with urethane (0.002 mg/kg), and their heads were stabilized using a specially designed holder. Differential recordings were made using pin electrodes (Grass Instruments, Quincy, MA). Two electrodes were implanted transcranially into the brain: a recording electrode was placed into the cerebellum just above the auditory brainstem nuclei, and a differential electrode was placed in the contralateral anterior cortex. A third (ground) electrode was inserted into the leg muscle. Responses were preamplified using a 10× stage pre-amp (A-M Systems, Carlsborg, WA), filtered (0.3-10 kHz bandpass), and amplified at 1000× (microelectrode amplifier model 1800; A-M Systems). Stimulus presentations for each frequency were begun at 120 dB and decreased in intensity in 10 dB steps in the range of amplitude that was well above threshold, and in 5 dB steps near threshold. Responses were averaged over 200 traces above threshold and 500 or 1000 traces near threshold. Threshold was defined operationally as the intensity at which the mean evoked response was twice the amplitude of the baseline electrical signal.

Subjective analyses of song

Subjective analyses of song were conducted as one means of assessing the effects of WN on song and of changes after removal of WN between D1 and D100. Two independent observers examined complete motifs of song and scored syllables as lost or changed. D1 songs were examined relative to PRE songs, whereas D100 songs were examined relative to D1 and PRE songs. Lost syllables simply did not appear in the later song sample, whereas changed syllables were typically in the same sequence in which they had been present and could be visually identified as exemplars of the same syllable despite gross changes in spectral properties. Changed syllables involved changes in fundamental frequency, harmonic structure, loss of specific subsyllabic structures, truncation (loss of notes within syllables), increased noisiness, or a combination of these. We used a conservative criterion for scoring syllables as changed such that only syllables that were quite visibly different between two time points were scored as changed; we did not attempt to assess subtle changes in syllable morphology. In addition, syllables were scored as changed or lost only if judged as such by both observers.

For the subset of syllables scored as changed at D100, two independent observers examined isolated exemplars of each and judged whether the D100 iteration was more similar to the D1 or the PRE version. This served as a preliminary test of whether syllables initially scored as obviously changed at D100 increased in similarity to their PRE iterations, providing possible evidence of song reacquisition. A second test was conducted in which the identity of the D1 and D100 syllables was unknown to the two observers. In the second analysis, all syllables present at all three time points were examined. Syllable renditions from D1 and D100 were compared with renditions from PRE, and each was scored for similarity on a scale of 1 (similar) to 3 (dissimilar). Syllables scored as 1 were judged as being no different from their PRE iterations, those scored as 2 were judged as being different from their PRE iterations but still fairly similar, and those scored as 3 were judged as being clearly dissimilar from their PRE iterations. Observers were aware neither of the syllable identity nor of the time point at which later recordings were made (PRE samples were labeled to enable comparisons). Scores for D1 and D100 were then compared to establish at which time point song was more similar to PRE, and only those instances in which both observers agreed were counted. This procedure allowed a direct comparison of the blind results with the initial analysis.

In all cases, sound analysis measures of syllable accuracy (described in more detail below) were generated to provide objective scores to supplement the subjective measures described above. Accuracy scores were used to verify the subjective classification of syllables as changed (Tables 1 and 4) and to verify subjective judgments of increases in similarity to PRE.

Analyses of temporal structure of song

Zebra finch song consists of a stereotyped sequence of three to eight discrete vocalizations called “syllables.” Syllables range in duration from ∼50 to 250 msec and are separated by 10-50 msec of silence. A sequence of syllables is called a “motif,” and multiple motifs are often sung in “bouts.” Zebra finches generally sing one canonical motif with little variation throughout their adult lives (Zann, 1996; Brainard and Doupe, 2001). Many birds also sing a limited number of “motif variants” (Williams and Mehta, 1997) wherein some subset of syllables is sung optionally. For example, a series of call-like notes may be sung between motifs in a bout but not at the end of a bout. Thus, there is normally some variability among songs with respect to temporal structure, although all birds produce one dominant motif. We used four measures to assess temporal structure of song and its change as a result of WN exposure: linearity, consistency, syllable probability, and motif probability.

To determine whether changes in any of these measures were the result of normal variation over time (Brainard and Doupe, 2001), we also made two sets of recordings from control birds 3 months apart and computed these statistics for both sets of recordings. One-tailed Mann-Whitney summed ranks tests with α levels of 0.05 were applied to comparisons between experimental groups and this set of control animals. One-tailed tests were justified because in all cases, we were expecting a greater degree of change in experimental than control groups. Nonparametric tests were used because of the small sample sizes and the noninterval nature of the measures.

Linearity. Linearity scores measure the degree to which syllables are sung in a specific order. A perfectly linear song is one in which syllable A is always followed by syllable B, syllable B is always followed by syllable C, and so forth. Thus, a nonlinear song is one in which syllables occur in different sequences in different renditions. Zebra finch songs tend to be highly linear because it is relatively rare for syllables to be sung in a variable order or to be repeated within a motif. However, the existence of motif variants may affect linearity if they result in syllables occurring consecutively on some occasions, but not others. For example, if a bird usually sings ABCDE, but CD is sometimes left out to form ABE, this will decrease the linearity score.

We used the following measure of linearity:

This measure differs somewhat from the linearity measure developed by Scharff and Nottebohm (1991) in that it does not count transitions at the ends of songs (Foster and Bottjer, 2001; Iyengar and Bottjer, 2002). This means that a perfect linearity score can be achieved even if a bird's song is occasionally truncated at the beginning and end. For example, if a bird sings ABC#BC#AB# (with pound signs representing the ends of motifs), it will receive a linearity score of 1 because there are two syllable-to-syllable transitions (AB and BC) and three syllables (A, B, and C). The same set of renditions would receive a score of 0.75 using the Scharff and Nottebohm (1991) measure because this measure includes the transitions C# and B# in its calculation of linearity.

Consistency. Consistency provides a slightly different measure of song stereotypy. A song can be linear in the sense that syllables always occur in the same order, but inconsistent in the sense that not all syllables are sung in all renditions. For example, if a bird sometimes sings ABCDE but occasionally sings ABC, this will affect the consistency score but not the linearity score. Furthermore, because linearity is computed by considering the absolute number of transitions, regardless of their frequency of occurrence, it is highly sensitive to any deviations from perfect linearity. In contrast, consistency measures the proportion of syllable transitions that are accounted for by the most frequent, or “dominant,” transition for a given syllable. This means that consistency will be less sensitive to infrequent deviations from a canonical pattern, while being more sensitive to deviations that do not affect linearity but are very frequent. In the example given above, for instance, if the bird sang the ABC variant only once, this would have only a moderate effect on consistency. In contrast, if the ABC variant were sung half the time and the ABCDE variant sung the other half, the consistency score would be quite low. Our measure of consistency was taken directly from Scharff and Nottebohm (1991):

In this measure, we count transitions (#) at both the beginnings and endings of motifs as syllables to capture variability at the beginnings and ends of songs. In the example given above, this measure gives a score of 0.83, reflecting the fact that the syllables #, B, and C participate in two different transitions; for example, for B, the following syllable could be C or #.

Syllable and motif probability scores. Williams and Mehta (1997) observed changes in the probability with which specific syllables and, as a result, specific song motif variants were sung after tracheosyringeal nerve sections (cf. Foster and Bottjer, 2001). We applied their measures of change in syllable probability and motif probability to examine the degree to which both individual syllables and specific song motifs were maintained after chronic exposure to WN.

Changes in syllable probability were defined as the mean difference between the probability with which syllables were sung before and after WN exposure. We computed the probability of each syllable being sung (i.e., the proportion of songs that included that syllable) both before WN exposure and at 1 d (D1) and 100 d (D100) after WN was discontinued. The mean absolute value of the difference between all PRE syllables and corresponding D1 syllables is the syllable probability change score for each bird, and it provides an overall measure of the amount of shift in the regularity with which song syllables were included in the song. Deleted syllables, as well as a decreased frequency of producing specific PRE syllables, can contribute to higher syllable probability scores.

Changes in motif probability were computed by determining the proportion of common motif variants present within each recording session both before and after WN exposure. That is, we identified the motif variant sung most often by each bird and added, in turn, the variants next most often sung until the list of variants accounted for at least 80% of the songs sung within each recording session. We then determined the percentage of original (i.e., PRE) common motif variants retained as common motif variants in subsequent D1 recordings for each bird as a measure of shifts in the composition of songs. Thus, a low score on this measure reflects a low carryover of initial motifs. We also compared D100 motifs to corresponding D1 motifs to examine recovery of song behavior after removal from WN.

Finally, for syllables that had been scored as changed at D100, two independent observers then examined isolated exemplars of each of those syllables and judged whether the D100 iteration of each syllable was more similar to the D1 or the PRE version. In all cases, qualitative assessments of syllable change at D1 and D100, and similarity of D100 syllables to pre-WN versus D1 iterations, were confirmed using sound analysis software to generate measures of syllable accuracy, which quantifies degree of similarity for individual syllables (see below and Tchernichovski et al., 2000).

Spectral analyses

We quantified overall spectral changes in song as well as in individual syllables using Sound Analysis software (Tchernichovski et al., 2000) to compare song behavior before WN (PRE), immediately after WN (D1), and after tutoring (D100). The software analyzes song along four separate dimensions (pitch, Wiener entropy, frequency-modulated noise, and spectral continuity) to arrive at a generic measure of similarity. This method of similarity analysis is superior to methods in which fast Fourier transforms of whole syllables are compared with one another via cross-correlation because it extracts dimensions that are relatively insensitive to normal variability in song production and recording (Tchernichovski et al., 2000). We also performed these analyses on the recordings from control birds made 3 months apart for comparison with experimental birds.

We used three different measures to characterize different aspects of change in song behavior: percentage similarity, accuracy, and spectral stereotypy. All spectral measures were computed by comparing 10 samples of song in each condition to 10 samples in another condition. Tests of significance for all spectral measures were performed using ANOVA with an α level of 0.05.

Percentage similarity. We analyzed overall spectral similarity of whole motifs using the “percentage similarity” measure to compare song samples at D1 and D100 to PRE songs. To generate this score for a pair of song samples, sound analysis compares the spectral characteristics of the first song sample to those of the second in 20 msec bins and then calculates the percentage of bins in the first song that exceeds the similarity threshold (default value, 92%) in the second, for a maximum percentage similarity of 100%. We compared PRE songs to D1 and D100 songs for each bird independently using the “overall” setting. We also compared samples of PRE song to other samples of PRE song as a control (note that this is the same as the “spectral stereotypy” measure below). At each time point, the dominant motif produced by each bird was used in these analyses.

Accuracy. Sound Analysis provides a second measure, “accuracy,” that we found to be better adapted for the analysis of individual syllables. Rather than compute the proportion of one sound that is similar to another sound, it reflects the degree to which two sounds are similar. For all parts of a pair of samples that exceed a given similarity threshold, the accuracy measure calculates a weighted average of the difference across the four dimensions analyzed. To ensure that as much of each syllable as possible was taken into account when computing accuracy, we set the similarity threshold to 0.15. Thus, for each pair of song samples, sound analysis generated an accuracy score, which was a weighted average of the similarity in pitch, entropy, frequency modulation, and continuity for all sections that exceeded the 0.15 similarity threshold. For each group, the mean accuracy score represents the mean for all syllables sung by birds in that group. Accuracy scores for D1 relative to PRE and for D100 relative to PRE were computed for all birds. To determine whether changes in accuracy as a result of WN exposure were within the range of normal variability, we compared these scores to scores generated by comparing samples of PRE song to other samples of PRE song as well as to measures from control subjects.

Spectral stereotypy. Finally, we used Sound Analysis to examine the variability in spectral properties of song within a given recording session, to generate what we call a spectral stereotypy measure. We measured spectral stereotypy by comparing randomly selected samples of song (whole motifs) within PRE, D1, and D100 recording sessions for experimental and control birds using the percentage similarity measure in sound analysis. This allowed us to observe changes in the stability of the spectral properties of song before and after WN and after tutor exposure. If a bird's song is highly stereotyped, renditions of individual syllables recorded at a given time point should be very similar to one another, and the spectral stereotypy score will be high. If there is a great deal of variability in the spectral properties of song, this score will be low.

Evoked potentials

As shown in Figure 1, auditory brainstem responses revealed a substantial shift in hearing threshold immediately after removal from WN. The mean threshold shift at the best frequency (3 kHz) was 27 dB 1 hr after removal from WN, although the general shape of the frequency sensitivity curve was preserved. Hearing showed substantial recovery by 3 d after removal from WN but was still ∼5-10 dB above that of normal controls. Hearing thresholds at best frequencies had recovered to normal after 1 week out of WN, although some deficits remained at very low (750 Hz and below) and very high (>4 kHz) frequencies.

Effects of WN exposure on song

Changes in song after removal from WN

After WN was discontinued, birds were exposed to a taped (group 1) or live (groups 2-4) tutor starting immediately after the D1 recording. Songs were recorded again at D100 and compared with D1 and PRE song. As noted above, there was substantial variability in the disruptive effect of WN on song that was not directly related to group membership. In addition, preliminary analyses suggested broad differences between severely and moderately disrupted birds in their recovery after restoration of auditory feedback. Therefore, we decided to analyze recovery as a function of degree of disruption measured immediately after removal from WN. We defined severely disrupted birds operationally as those for whom half or more of all song syllables were lost or changed as a result of WN exposure (n = 9 on D1). Moderately disrupted birds were those for whom fewer than half of all song syllables were lost or changed (n = 13 on D1).

To validate the categorization of birds into “severe” and “moderate” groups, we performed spectral analyses for birds in the two groups by comparing their PRE song behavior to that recorded on D1. As shown in Table 3 (see also Fig. 4), the severe group was more disrupted than the moderate group, as judged by all three measures used: percentage similarity (F_(1,20) = 7.03; p < 0.01), accuracy (F_(1,19) = 5.10; p < 0.05), and spectral stereotypy (F_(1,20) = 8.04; p < 0.01). Interestingly, categorizing birds in this manner resulted in a particularly pronounced difference in the incidence of syllable loss, as opposed to syllable change: severely disrupted birds lost a total of 18 PRE syllables from their D1 songs, whereas moderately disrupted birds lost only one syllable; in contrast, severely disrupted birds changed a total of 23 of their syllables compared with 14 in moderately disrupted birds. Thus, severe disruption of song because of removal of auditory feedback seems to entail a disproportionate loss of syllables from stable vocal patterns.

Effects of WN are similar to deafening

Long-term exposure to WN resulted in changes in spectral and temporal aspects of adult song behavior similar to those observed in surgical deafening studies (Nordeen and Nordeen, 1992; Brainard and Doupe, 2001). In the most severe cases, a majority of the syllables sung before WN exposure were lost or radically changed. Although song sequence changed as a result of lost syllables, changes in the ordering of syllables were much smaller in magnitude than changes in syllable phonology. Birds varied considerably in the severity of disruptions induced by WN. Interestingly, the age at which our subjects were exposed to WN corresponds to the time point at which the most variability in the effect of deafening has been observed in studies examining the influence of age at deafening on the rate and severity of song disruption (Lombardino and Nottebohm, 2000; Brainard and Doupe, 2001).

Changes in song after restoration of normal auditory feedback

Song change after removal from WN was highly variable. Although certain aspects of song were reacquired after removal from WN, recovery of original song patterns was incomplete, especially in birds with severely disrupted songs. For example, approximately half the syllables that changed as a result of exposure to WN did not improve by 100 d after noise removal. Furthermore, reacquisition of individual syllables was often partial and never resulted in complete recovery of a severely disrupted song.

The reacquisition of syllables that had been disrupted, and the overall increase in similarity to PRE song in some moderate subjects (Table 5; Fig. 6), suggests that some form of sensorimotor template is maintained in adult birds. However, it is also possible that some of these changes were the result of adjustments of fairly general parameters of song (Fee et al., 1998). For example, the deletion of the downsweep in the syllable from subject y665 on D1 (Fig. 6) could have resulted from decreased air flow alone (i.e., the syrinx may still have been producing the same gesture as in PRE and D100 song). If this were true, the reacquisition of this note would not require any new learning but only an increase in airflow between D1 and D100. This seems less likely in the case of the syllable from subject pu781 (Fig. 6); in this case, changes occurred at a number of levels, including not just general properties such as source quality and the bandwidth of harmonics, but the overall structure of the syllable as well, from downsweep-stack-downsweep to downsweep-stack-stack (for methods of classifying syllables, see Zann, 1996).

Failures to reacquire initial song structure also provide some insight into the basic mechanisms that underlie learning in the song system. For example, increases in song stereotypy that do not reflect reacquisition of song after restoration of feedback may be guided by a generic species-specific template (Marler and Peters, 1982) in addition to (or instead of) a learned template. The data are also consistent with the notion that song learning is achieved by a reinforcement mechanism, as suggested by Troyer and Doupe (2001a,b; Dave and Margoliash, 2001). Reinforcement learning proceeds by strengthening synaptic connections that lead to an outcome consistent with some target. As such, it requires that the target be approachable by successive approximation given the variability in the behavior itself. In severely disrupted birds, song output may deviate enough from the target stored in the template that this target cannot be reacquired. Instead, the reinforcement signal leads to a stabilization of highly abnormal song.

Another possibility is that the template itself remains susceptible to change even in adulthood. For example, the fact that lost syllables were never recovered could indicate that such syllables were actively “unlearned” as a result of disruption of auditory feedback, perhaps by being eliminated from the bird's initially acquired template. Mechanisms for active template changes in adult birds based on usage or experience could be similar to those used during initial song development; for example, Solis and Doupe (1997) found that responses in lateral magnocellular nucleus of the anterior neostriatum (lMAN) evolved during development to be selective to a bird's own song, even when this song was manipulated to be dissimilar from the putative template song.

Temporal structure and spectral properties of syllables exhibit plasticity under different circumstances

As indicated above, the failure to reacquire lost syllables may indicate elimination of syllables from the template. This pattern is consistent with results from an experiment by Hough and Volman (2002) in which beads were implanted into the syrinx, resulting in the temporary distortion of some syllables and silencing of others. In some instances, the interval between syllables adjacent to the silenced syllable was shortened. The only permanent effect of bead implants was the permanent deletion (even after bead removal) of omitted syllables for which this silent “gap” had been closed. This observation is also consistent with a key finding of Tchernichovski et al. (2001), who studied song acquisition in juvenile birds exposed to a highly efficient tutoring regimen: when a syllable at a given position in the tutor's song was similar to a syllable at a different position in the juvenile's song, the juvenile would change a syllable in the same position rather than move the similar syllable to the appropriate position.

Thus, it may be that the temporal structure of song is less plastic under most circumstances than individual syllable phonology. This would also explain why, leaving aside changes resulting from the loss of individual syllables, temporal changes observed as a result of WN exposure in the current work were minimal and rarely resulted in the reordering of song elements. Furthermore, even disrupted songs maintained a high degree of linearity and consistency, suggesting that they were sung in a stereotyped sequence despite gross spectral disruptions.

Data from the delayed auditory feedback study of Leonardo and Konishi (1999) suggest that there are circumstances under which the temporal structure of song may be more plastic than spectral structure. Their procedure left normal auditory feedback intact, such that birds heard a superposition of normal and introduced sounds. This procedure produced substantial disruption of the temporal organization of song, although birds continued to produce a low incidence of their normal songs even when overall song disruption was maximal. After removal of delayed auditory feedback, birds in the study by Leonardo and Konishi (1999) gradually recovered their previous song patterns.

The differential plasticity of syllable phonology and higher-order temporal structure may arise because these two parameters of song are associated with anatomically distinct neural substrates. One possibility is that high vocal center (HVC) codes primarily for higher-order temporal structure, whereas robust nucleus of the archistriatum (RA) codes for syllable phonology (Vu et al., 1994; Yu and Margoliash, 1999; Hahnloser et al., 2002), and these nuclei lie on different pathways with respect to the MAN-basal ganglia circuit that supports song learning. Whereas RA receives direct input from lMAN, HVC receives input from medial magnocellular nucleus of the anterior neostriatum (mMAN). Foster and Bottjer (2001) demonstrated that lesions to mMAN result in selective disruption of temporal aspects of song, particularly at the beginnings of song bouts.

In contrast to Bengalese finches, zebra finches do not learn new song elements

Woolley and Rubel (2002) found that some syllables in the songs of Bengalese finches reversibly deafened with ototoxic hair cell lesions gradually evolved to resemble those of birds with which they were housed during recovery, whereas we found no evidence for learning from live or taped tutors in adult zebra finches. Bengalese finches sing more complex songs than zebra finches (they repeat each syllable a variable number of times from rendition to rendition, whereas zebra finches typically sing each syllable once per motif) and have much higher rates of neurogenesis (at least after deafening) (Scott et al., 2000). Bengalese finch song may therefore be more plastic than zebra finch song, as evidenced by the relatively short period of deafening required to induce song deficits in Bengalese finches (cf. Nordeen and Nordeen, 1992, with Woolley and Rubel, 1997). Of course, we must be cautious in concluding that the difference between our results and those of Woolley and Rubel (2002) are attributable solely to species differences in plasticity. Whereas we exposed our subjects to WN, their subjects were deafened by cochlear lesion, which could promote injury-induced plasticity in auditory neural circuitry.