Lower SNR sound event recognition using noisy training sample

Sound event recognition becomes a basic task in some applications. However, in low SNR condition, the accuracy rate is easily affected by the acoustic scene. To address the problem, this paper proposes a framework consisting of empirical mode decomposition (EMD), gray level co-occurrence matrix combined with higher-order singular value decomposition (GLCM-HOSVD), and random forests (RF). We use a voting method based on the first to sixth intrinsic mode functions (IMFs) which is generated from EMD, to detect the endpoint of sound events and estimate the SNR. GLCM-HOSVD is proposed to extract features from audio data. During classifier training, the sound samples mixed by environmental sound and sound events are used to train RF. The experiment proves that the proposed method has the ability to recognize low SNR sound events in acoustic scenes. The result shows that the accuracy rate is higher than 78% even in 5dB acoustic scene.