A Dynamical Model of Pitch Memory Provides an Improved Basis for Implied Harmony Estimation

Area of Science:

• Music Information Retrieval
• Computational Auditory Neuroscience
• Signal Processing

Background:

• Automatic chord estimation traditionally relies on chroma features, which struggle with unaccompanied melodies containing non-chord tones.
• Accurate estimation of implied harmony from tonal melodies remains a challenge.

Purpose of the Study:

• To develop a novel approach for automatic chord estimation by modeling human perception of pitch sequences.
• To differentiate chord tones from non-chord tones in tonal melodies using principles of auditory short-term memory.

Main Methods:

• A gradient frequency neural network, modeling auditory short-term pitch memory, was employed.
• The model simulates pitch salience through nonlinear resonance and lateral inhibition among tonotopically tuned oscillators.
• The model was tested as a feature extractor on unaccompanied tonal melodies.

Main Results:

• The model selectively enhances chord tones, leading to improved accuracy in implied harmony estimation.
• Performance gains were observed when using segmented input signals, similar to existing methods.
• The model demonstrates potential as a biologically realistic feature extractor for chord estimation.

Conclusions:

• The proposed neural network model, based on human pitch perception, offers a promising new direction for automatic chord estimation.
• This approach effectively addresses limitations of chroma-based methods, particularly for unaccompanied melodies.
• Further development within a dynamical systems framework could lead to a complete chord estimation system.