A tonal features exploration algorithm with independent component analysis

In this paper, the tonal feature exploration (TFE) algorithm is addressed to intent to explore representative audio features, which can offer rather discriminative power to meet the classification of human perception. First of all, TFE algorithm searches the defined tonal tracks from the successive overlapped frames by tracing and linking the near audio tones. The tonal track plays the masking role of specific segments where the representative frames are fetched along the tonal tracks based on their significance. The unmasked tonal energies for different subbands resulting from the audio psychoacoustic module are aggregated and arranged into the observation data matrix. Finally, the signature projection matrix can be learned through performing the independent component analysis (ICA) process on the observation matrices that the components of matrix are quantized as the index of an audio sequence. The precision and recall rates of the simulation can demonstrate that the TFE algorithm can extract the concise audio features for the retrieval of audio content-based database.