Noise-invariant Neurons in the Avian Auditory Cortex: Hearing the Song in Noise
A. Performance of three noise reduction algorithms (STRF, OWF, SINR) and lower and upper bounds (Stim, IBM) on song embedded in colony noise or modulation-limited (ML) noise. The performance ratio (y-axis) depicts the improvement in noise levels over the noise-corrupted signal, as measured by the cross-correlation in the log spectrogram domain, with the error bars representing one standard deviation across five noisy stimuli. On the x-axis are the models we have tested, where “Stim” is the noise-corrupted signal, “STRF” is the model presented here, “OWF” is the optimal Wiener filter, “SINR” is a spectral subtraction algorithm similar to STRF but based on engineering constructs, and “IBM” is an ideal binary mask. B. Spectrograms of the signal masked with noise from the zebra finch colony, the clean zebra finch song, and our signal reconstruction, followed by the time-frequency gains. C. Same as B but for modulation-limited noise. Sounds can be found in the SI.