Key-finding by artificial neural networks that learn about key profiles.

Area of Science:

• Music cognition
• Computational musicology
• Artificial intelligence

Background:

• Determining the musical key of auditory stimuli is a fundamental task in music cognition.
• Traditional algorithms exist for key-finding but may lack nuanced feature weighting.

Purpose of the Study:

• To investigate the efficacy of a simple artificial neural network, a perceptron, in identifying musical keys.
• To compare the performance and internal mechanisms of perceptrons against established key-finding algorithms.

Main Methods:

• Perceptrons were trained using standardized musical key profiles from various sources.
• Trained perceptrons were evaluated on their accuracy in asserting the musical key of 296 novel auditory stimuli.
• Analysis included comparing perceptron activity patterns and weight assignments to traditional algorithms.

Main Results:

• Perceptron accuracy in key assertion was comparable to established algorithms, varying with training data.
• Perceptrons exhibited distinct activation patterns, favoring selected keys and suppressing competing keys more strongly than traditional methods.
• Internal analysis revealed that perceptrons assign differential weights to key profile components, identifying more salient features.

Conclusions:

• Simple perceptrons demonstrate a viable approach for musical key assertion, achieving performance on par with existing methods.
• Perceptrons offer a unique mechanism by learning feature importance, which could enhance traditional key-finding algorithms.
• The differential weighting strategy learned by perceptrons presents a novel avenue for improving computational music analysis.