Robust query-by-singing/humming system against background noise environments

Abstract

Under background noise environments, the performance of the Query-by-Singing/Humming (QbSH) system is considerably degraded. Since human pitch information is used as a feature vector for the QbSH system, a noise robust pitch estimation algorithm is inevitable. Thus, a novel pitch-estimation method is proposed by integrating temporal-autocorrelation and spectral-salience methods. As a pre-processing block, spectral smoothing is applied to enhance the stationarity of the noisy input signal. To calculate the similarity between the MIDI database and input humming signal, the dynamic time warping (DTW) algorithm is used. Jang's corpus and AURORA2 database are selected as humming and background noise signals, respectively. Compared with the standard pitch estimation algorithm in the ITU-T G.729 speech codec, the proposed pitch estimation method improves the average accuracy by 11.7% for the 0 dB signal-to-noise ratio (SNR) noise case. It also improves top-20 ratio and mean reciprocal rank (MRR) of the proposed QbSH system, on average, by 7.4% and 0.13, respectively.