Auditory Feature-based Perceptual Distance.

Area of Science:

• Bioacoustics
• Computational Neuroscience
• Machine Learning

Background:

• Traditional methods for comparing acoustic signals, such as mean squared error (MSE), rely on pixel-wise differences in spectrograms.
• These pixel-wise metrics often fail to capture perceptual sensitivity, being overly sensitive to minor, imperceptible signal variations.

Approach:

• We propose the auditory perceptual distance (APD) metric, utilizing acoustic features extracted by a convolutional neural network (CNN).
• A Siamese CNN was trained on songbird vocalizations, with spectrograms rescaled to mimic European starling auditory frequency sensitivity.
• APD is defined as the cosine distance between feature vectors from the trained CNN, offering a more perceptually relevant comparison.

Key Points:

• The APD metric demonstrates greater robustness against temporal and spectral translations compared to MSE.
• APD accurately models the sigmoidal relationship observed in behavioral psychometric functions across complex acoustic spaces.
• Fine-tuning APD with starling behavioral data improved predictions of perceptual sensitivity, discrimination, and categorization for novel acoustic stimuli.

Conclusions:

• The APD metric significantly outperforms MSE in both robustness and perceptual accuracy for analyzing acoustic signals.
• APD's tunability allows it to adapt to experience-dependent perceptual biases, making it a versatile tool for bioacoustic research.
• This approach offers a more biologically plausible and accurate method for quantifying differences in complex acoustic signals.