Noise Robust Music Artist Recognition Using I-Vector Features

In music information retrieval (MIR), dealing with different types of noise is important and the MIR models are frequently used in noisy environments such as live performances. Recently, i-vector features have shown great promise for some major tasks in MIR, such as music similarity and artist recognition. In this paper, we introduce a novel noise-robust music artist recognition system using i-vector features. Our method uses a short sample of noise to learn the parameters of noise, then using a Maximum A Postriori (MAP) estimation it estimates clean i-vectors given noisy i-vectors. We examine the performance of multiple systems confronted with different kinds of additive noise in a clean training noisy testing scenario. Using open-source tools, we have synthesized 12 different noisy versions from a standard 20-class music artist recognition dataset encountered with 4 different kinds of additive noise with 3 different Signal-to-Noise-Ratio (SNR). Using these datasets, we carried out music artist recognition experiments comparing the proposed method with the state-ofthe-art. The results suggest that the proposed method outperforms the state-of-the-art.