Multi-channel audio signal retrieval based on multi-factor data mining with tensor decomposition

Abstract

In this paper, an efficient method of multi-channel audio signal retrieval with finite number of channels is proposed based on multi-factor data mining with tensor decomposition. We briefly discuss how to convert the limited channels to an increased number of channels (multi-channel) by capturing the latent higher-order tensor structure of multi-channel audio data. The multi-channel audio data space is established mainly due to three factors including location, channel and time-frequency. Moreover, CANDECOMP/PARAFAC (CP) decomposition is introduced in the process of multi-factor data mining to predict the data in the missing channels. Besides, considering human auditory effects at low frequency, we compute a set of data in advance for the retrieval of Low Frequency Effects (LFE) channel. The performance of the proposed method is assessed by MUlti-Stimulus test with Hidden References and Anchor listening test (MUSHRA). We further demonstrate the retrieval of 5.1 multi-channel audio from stereo audio. Experiments show that an acceptable converting quality has been achieved and the novel tensor-based method is easy to implement as compared to the traditional method.