221 / 1971-01-01 00:00:00

Short-term Specific Audio Detection

5567

Abstract—There are several difficulties in short-term specific audio detection: firstly, the input has less information and it is very intractable to extract suitable feature for representing the signal because of the short duration of the audio. Secondly, a problem of the shortage of training data occurs if directly using the statistic model to process the short-term specific audio. Thirdly, the difficulty of detection increases rapidly when the background noise is relatively large and complex. In this paper, an algorithm based on universal background model-Gaussian mixture model (UBM-GMM) for short-term specific audio detection is proposed, which not only overcomes the defects of lack of data but also in a certain extent can suppress the noise. In this paper, experiments in which there are three kinds of specific audio with different signal-to-noise rate (SNR) are done, the results show that the algorithm can detect the short-term specific audio well in different SNR.