Visually-guided compensation of deafening-induced song deterioration

Abstract

Human language learning and maintenance depend primarily on auditory feedback but are also shaped by other sensory modalities. Individuals who become deaf after learning to speak (post-lingual deafness) experience a gradual decline in their language abilities. A similar process occurs in songbirds, where deafness leads to progressive song deterioration. However, songbirds can modify their songs using non-auditory cues, challenging the prevailing assumption that auditory feedback is essential for vocal control. In this study, we investigated whether deafened birds could use visual cues to prevent or limit song deterioration. We developed a new metric for assessing syllable deterioration called the spectrogram divergence score. We then trained deafened birds in a behavioral task where the spectrogram divergence score of a target syllable was computed in real-time, triggering a contingent visual stimulus based on the score. Birds exposed to the contingent visual stimulus-a brief light extinction-showed more stable song syllables than birds that received either no light extinction or randomly triggered light extinction. Notably, this effect was specific to the targeted syllable and did not influence other syllables. This study demonstrates that deafness-induced song deterioration in birds can be partially mitigated with visual cues.

Keywords: deafening; degradation; sensorimotor; sensory feedback; songbird; vocal control.

Figure 1

A new feature to assess sound degradation. (A) Spectrogram of three consecutive motifs produced by a single bird on a baseline day. An algorithm is trained to detect a song syllable (black arrows) and to extract its spectrogram from the 48 ms window before the detection point (dashed red box). The median of all 48-ms syllable excerpts provided the “Reference spectrogram.” On the following days, the Euclidean distance was computed between the spectrograms of the targeted syllable and the Reference spectrogram to obtain the Spectrogram divergence score. (B) Waveforms (top) and spectrograms (bottom) of the same song motif from a single bird (labeled b10o8) with increasing levels of white noise (WN) added to the signal and of pure white noise. The dashed red box highlights this bird’s 48-ms syllable excerpt used to assess the spectrogram divergence score. (Ci) Violin plot showing the spectrogram divergence score (left y axis) computed between 200 randomly selected renditions of the 48-ms syllable excerpt for bird b10o8 with increasing levels of white noise in the signal (see B for spectrogram representation) and its reference spectrogram computed in the absence of white noise. White dots on the violin plots correspond to the mean spectrogram divergence scores. Note that, whatever the level of WN, there were no errors in the computation of the spectrogram divergence score (right y axis). (ii) Euclidean distances between 200 randomly selected song motifs with the same motifs that include increasing levels of white noise in the signal or with pure white noise (left y axis). There is again no error in the computation of the score (right y axis). (iii) Violin plots showing the dissimilarity score computed using Sound Analysis Pro (SAP; formula: 100-similarity score) between 200 randomly selected song motifs with the same motifs that include increasing levels of white noise in the signal or with pure white noise (left y axis). The percentage of errors of computation (when SAP failed in computing the similarity score between two songs) increases with the level of white noise added to the signal (green bar plot, right y axis). (iv) Violin plot of the entropy between 200 randomly selected song motifs with the same motifs that include increasing levels of white noise in the signal or with pure white noise. (D) Spectrogram divergence scores (i), Euclidean distance (ii), SAP dissimilarity score (iii) and SAP entropy (iv) computed as in (C) for 19 birds. Mean is shown with a thick black line and individual traces are in grey. Example bird from (A–C) (b10o8) is shown in red.

Figure 2

A behavioral task to counteract deafening induced song degradation. (A) Birds were divided into four groups. All birds were isolated in a sound-proof chamber and their song was recorded on a few consecutive baseline days. Except the hearing birds (n = 7 birds), all the other birds were deafened following the procedure described in (Zai et al., 2020). After the deafening procedure, birds recovered for a few days before the onset of the behavioral paradigm. (B) During the task, songs were recorded online and a target syllable specific for each bird was automatically detected. For deaf LO birds (n = 13 birds), whenever the spectrogram divergence score was lower than a certain threshold, the housing light in the sound-proof chamber was transiently switched off for 200 ms, the light remained on otherwise. For deaf no LO birds (n = 15 birds), the light was never switched off while for deaf random LO birds (n = 5 birds), the light was transiently but randomly switched off when the bird produced the target syllable. (C) Spectrograms of song motifs produced by three birds from the same clutch before and 1 to 6 weeks post-deafening and onset of the LO protocol. The target syllable for each bird is shown with a dashed orange box and zoomed in versions produced pre-deafening and 6-weeks post-deafening are highlighted. (D) UMAP projections of all sounds recorded (i.e., birds vocalizations and cage noise) in the sound proof chamber for each example deafened bird and for a hearing control bird before deafening (blue) and 6 weeks post-deafening and onset of LO protocol for deafened birds, or 7 weeks later for the hearing bird. Clusters that include bird vocalizations (songs and calls) and noise are surrounded by dashed red and blue lines, respectively.

Figure 3

Transient light extinction contingent on the spectrogram divergence score slows syllable degradation in deaf birds. (A) Variations in spectrogram divergence scores (black dots, left y axis) and the number of detected target syllables (blue line, right y axis) over days for three example deafened birds (same birds as in Figure 2): the deafened 11C no LO (left), the deafened 11E LO (middle) and the deafened 11R random LO bird (right). Black line indicates the mean value of the spectrogram divergence score with the dots representing the raw values calculated for each detected target syllable. Red line for the deafened 11E LO bird indicates the threshold value used for triggering the transient light extinction. The symbol in orange indicates when the deafening procedure occurred. The LO exposure started on day zero, with the same day used for both the 11C and the 11E birds. Inserts show the mean (+/-SEM) spectrogram divergence score over the last day before the onset of the LO protocol and days 25–30 after the onset of the LO protocol. Shown above is the percent of change. (B) Normalized spectrogram divergence scores (mean Z-score+/-SEM) computed on all detected renditions of the 48-ms target syllable for all birds before and during the LO protocol. (C,D) Entropy (C) and entropy variance (D) computed on a subset of 200 randomly selected renditions per week per bird of the 48-ms target syllable. (E) The number of target syllables automatically detected decreased for all deafened birds. *, significant group effect, p < 0.05 (see text for details).

Figure 4

Various acoustic features show no clear evidence of an impact of a LO exposure on the structure of the entire target syllables or of other song syllables. Acoustic measures were computed on subsets of 200 randomly selected target (left panel) and non-target (right panel) syllables per bird and per week. Spectrogram divergence score (A), syllable duration (B), entropy (C) and entropy variance (D) of the entire target (i) or non-target (ii) syllables over the 6 weeks after the onset of the LO protocol for deaf LO birds. Spectrogram divergence scores evolved within a similar range for the three groups of deafened birds, deaf LO (blue), deaf no LO (red) and deaf random LO (purple). *, significant group effect, p < 0.05 (see text for details).