Audio Type Detection Method Based on Bipolar Modeling of Pure Speech and Background Noise

Abstract

The present invention provides an audio type detection method based on pure speech and background noise bipolar modeling, comprising the steps of: S1, constructing a pure speech GMM model and a pure background noise GMM model; Distance, to judge whether the Gaussian mixed element is located in the feature overlapping space; S3, remove the Gaussian mixed element located in the feature overlapping space, rebuild the pure speech statistical model, the pure background noise statistical model; S4, calculate the new pure speech statistical model, pure background noise The probability of the noise statistical model, and the estimated signal-to-noise ratio of the audio sample; S5, according to the calculated probability and estimated signal-to-noise ratio, construct the feature vector and use the SVM model to make a judgment, and distinguish the sample as pure speech, background noise or noise-containing speech . The invention can effectively distinguish pure speech, pure background noise and noise-containing speech while reducing the calculation amount of GMM-SVM.

Description

technical field [0001] The invention relates to the technical field of speech signal processing, in particular to an audio type detection method based on pure speech and background noise bipolar modeling. Background technique [0002] Audio type detection technology refers to the use of different types of audio features to distinguish their types. At present, the technologies used more frequently are GMM (Gaussian Mixture Model) model and HMM (hidden Markov model) model. In recent years, the combination method of GMM-SVM has appeared, using the GMM model to construct supervectors as features, and using the SVM (Support Vector Machine, Support Vector Machine) model for soft classification, and good results have been obtained. The calculation amount of GMM-SVM is determined by the Gaussian mixture degree of GMM, the supervector dimension used and the SVM kernel function. The supervector dimension is generally the feature dimension or Gaussian mixture degree used by GMM, and t...

AI Technical Summary

Problems solved by technology

Therefore, GMM-SVM has a large amount of computation. For applications that require real-time processing, there are certain performance requirements for the equipment.

Images