Interestingly, neurons in each brain area were equally good at extracting trained and unfamiliar songs (data not shown), indicating that training and behavioral relevance were not critical for the neural extraction of songs from auditory scenes. Furthermore, segregating neurons BAY 73-4506 mw in the midbrain and primary AC into broad and narrow populations revealed no significant differences in the extraction of songs from auditory scenes (Figure S4). These findings show that BS neurons represent individual songs in auditory scenes at SNRs that match birds’ perceptual abilities to recognize songs in auditory scenes, in contrast to NS and see more upstream neurons.

The BS population represented individual songs with a sparse population code, in contrast to the representation of songs in upstream populations, and we next aimed to understand how a sparse sensory representation arises in the BS population. One neural mechanism for producing sparse sensory responses is with neurons

that are only sensitive to very specific stimulus features. To determine whether BS neurons were sensitive to particular acoustic features, we computed a percentage similarity score (Sound Analysis Pro, Tchernichovski et al., 2000) for every pair of notes to which an individual BS neuron responded. Percentage similarity score describes the acoustic similarity of a pair of notes based on measures of pitch, amplitude modulation, frequency modulation, Weiner entropy, and goodness of pitch. Like neurons in other auditory populations, pairs of notes to which a BS neuron responded were spectrotemporally more similar to one another (percentage similarity score, much 69.2 ± 28.3) than were notes selected at random (percentage similarity score, 45.8 ± 27.2, mean ± SD; p < 0.0001; Figures S6A and 6B). However, unlike other recorded neurons, BS neurons often failed to respond to every iteration of a note that was repeated multiple

times in a song (Figure S6C; see Figure 7A), and notes that were spectrotemporally similar to a response-evoking note often failed to evoke a response (see Figure S6B). These observations indicate that although individual BS neurons were sensitive to particular acoustic features, acoustic features alone may be inadequate for predicting their responses. To quantitatively assess the acoustic features to which BS neurons were tuned, we next computed spectrotemporal receptive fields (STRFs). STRFs provide an estimate of the acoustic features to which a neuron is sensitive, and the complexity of a receptive field can indicate a neuron’s selectivity for complex or rarely occurring acoustic features.