Note and Timbre Classification by Local Features of Spectrogram

Abstract

In recent years, very large scale online music databases containing more than 10 million tracks became prevalent as the fostered availability of streaming and downloading services via the World-Wide Web. The set of access schemes, or Music Information Retrieval (MIR), still poses several and partially solved problems, especially the personalization of the access, such as query by humming, melody, mood, style, genre, instrument, etc. Generally the previous approaches utilized the spectral features of the music track and extracted several high-level features such as pitch, cepstral coefficients, power, and the time-domain features such as onset, tempo, etc. In this work, however, the low-level local features of the spectrogram partitioned by means of the Bark scale are utilized to extract the quantized time-frequency-power features to be used by a Support Vector Machine to classify the notes (melody) and the timbre (instrument) of 128 instruments of General Midi standard. A database of 3-second sound clips of notes C4 to C5 on 7 sound cards using two software synthesizers is constructed and used for experimental note and timbre classification. The preliminary results of 13-category music note and 16-category timbre classifications are promising and their performance scores are surpassing the previously proposed methods.