Endpoint Detection and De-noising Method Based on Multi-resolution Spectrogram

Abstract

The paper studied endpoint detection algorithm of noisy speech, since the visual differences of spectrogram employed by speech and noise, the paper chose spectrogram endpoint detection methods. Technical difficulties of spectrogram endpoint detection is how to describe the intuitive difference of spectrogram by mathematical amount, according to the descriptive power of autocorrelation coefficients on texture features, the paper described the difference by selecting the autocorrelation function, and proposed column autocorrelation spectrogram detection method. Through the distribution of spectrogram self-correlation function, as the threshold of endpoint detection for the noisy speech, the cut-off point between speech and noise was found out. Since the paper used broadband spectrogram, which employed poor frequency resolution, so there were still residual noise in speech column after autocorrelation spectrum detection, in order to further de-noising in different bands, combined with the multi resolution of empirical mode decomposition (EMD), the paper analyzed the noisy speech by multi-resolution, the target was broken down into different frequency scales and was further analyzed by column autocorrelation spectrogram, experiments shown that the noise reduction effect for noisy speech was ideal.

This work is supported by Guangdong Provincial Science and technology projects#2013B040401015.