Noise spectrum estimation and voice mobility detection method based on unsupervised learning

Abstract

The invention relates to a noise spectrum estimation and voice mobility detection method based on unsupervised learning, which comprises the steps of: 1, establishing a GMM (Gaussian Mixture Model) model aiming at logarithm amplitude features of a voice signal on each frequency point; 2, setting M frames of buffers for one section of voice data, storing the former M frames of input signals into the buffers, extracting a logarithm amplitude spectrum of M frames in the buffers, and substituting into the GMM mode of the step 1 for initializing to obtain an initialized model Lambda0,k; and 3, updating the GMM model by frames by adopting an incremental learning mode from the (M+1)th frame after the initialize model Lambda0,k is obtained, and carrying out sequential recursion to obtain chances of occurrence of the noise value and the voice signals on the kth frequency point of the ith frame. The invention is a tight coupling solution of spectrum estimation and voice mobility detection, which can enhance the adaptability of the voice application system to the noise environment. The invention is independent from the hypothesis of the noise initialization, and can provide the description of the voice mobility on the time frequency two-dimensional space.

Description

technical field The present invention relates to the technical field of speech signal processing, in particular, the present invention relates to a noise power spectrum estimation and speech activity detection method based on unsupervised learning. Among them, speech activity detection is an algorithm for judging the presence or absence of speech in the time dimension. It can not only answer the existence of speech in the form of "yes" or "no", but also describe the presence of speech with the probability of speech occurrence. Background technique Most speech application systems have to deal with environmental noise interference. Predecessors have proposed many methods to remove the interference of noise on the speech system, almost all of which rely on speech activity detection and noise power spectrum estimation. There is a close relationship between these two modules, and their accuracy directly affects the overall noise immunity of the system. Traditional solutions hav...

AI Technical Summary

Problems solved by technology

Aiming at the shortcomings of the previous voice activity detector and noise power spectrum estimator, the present invention proposes a tightly coupled solution, so that voice activity detection and noise power spectrum estimation are unified under an unsupervised learning framework, thereby enhancing Adaptability of Speech Application System to Noise Environment

Images