Robust audio fingerprinting based on GammaChirp frequency cepstral coefficients and chroma

Abstract

A novel auditory feature that combines an auditory model and music theory is proposed for audio fingerprinting. First, the input audio is filtered by a GammaChirp (GC) filterbank to model the cochlear frequency selectivity. Then, the output of the filterbank is down-sampled and decorrelated by a discrete cosine transform to obtain the GammaChirp frequency cepstral coefficients (GCFCCs). Next, some lowest order GCFCCs are projected onto the chroma to characterise both melodic and harmonic information of the input. Finally, non-negative matrix factorisation is applied to the chroma matrix to reduce its dimension while maintaining its discriminative power. The experimental results illustrate that the proposed scheme achieves a stabler identification rate and lower computational complexity than the schemes based on the Mel-frequency cepstral coefficients.