Timbre Characterization with Mel-Cepstrum and Neural Nets

In this work the problem of timbre recognition--classification is addressed by com-bining the properties of a powerful speech-coding technique, the Mel-frequency Cepstral Coefficients, with the feature extraction capabilities of a self-organizing neural network. Acoustic relationships between tones are reflected into spatial relationships onto a neural lattice. Final results are in good agreement with the usual classifications of timbre quality, and offer promising grounds for the con¬struction of a general, analysis-based timbre space.