Novel dual-microphone speech detection and enhancement method

Abstract

The invention relates to the field of speech information processing technology and microphone array signal processing and especially relates to the fields of voice activity detection, voice detectionand speech recognition and interaction and the like. The method can effectively detect time frame of a voice activity and can also carry out dynamic regulation on threshold according to noise transform through two dynamic threshold update policies, by taking variability of noise environment into full consideration and based on three groups of auditory features capable of reflecting the ratio of noise energy in the total energy. The detection result can be corrected again through a detection result buffer mode, thereby preventing the defect of detection miss between continuous active voice frames; a noise power spectral density matrix is subjected to adaptive update according to the speech detection result; and furthermore, speech enhancement can be carried out through a Weiner filter, so that the noise can be suppressed under the minimum mean square error criterion.

Description

technical field [0001] The invention relates to the field of speech recognition and detection, in particular to a method for double-microphone speech detection and enhancement based on a dynamic threshold updating strategy. Background technique [0002] Affected by machine learning technologies such as deep neural networks, the accuracy of speech recognition has been greatly improved, and speech recognition has begun to be widely used in various fields. At present, speech recognition technology is mostly used in various electronic devices such as mobile phones, air conditioners, and TVs. Compared with traditional remote controls, the human-computer interaction technology of speech recognition is more convenient, and it is a new generation of information query and information recommendation without interactive interface. The key to human-computer interaction technology. [0003] At present, in the absence of strong noise interference and near-talk, the accuracy of speech rec...

AI Technical Summary

Problems solved by technology

[0004] There are currently three methods for automatic speech detection, which are the short-term energy in the time domain, the zero-crossing rate, and the mean square of frequency-band energy in the frequency domain. Then it is compared with an empirical threshold. Practical applications show that these three methods have two main disadvantages: 1. With a fixed threshold, it is impossible to achieve good performance when the noise environment changes, and in practical applications the noise is usually Changeable, it is difficult to find a suitable fixed threshold to adapt to most noise scenarios; 2. The method of comparing the short-term energy or zero-crossing rate alone has unstable performance and low accuracy when the noise energy is strong. At the same time, if the accuracy of voice detection is low, statistical information such as noise power spectrum will be inaccurate, or voice information will be included by mistake, resulting in voice distortion

Images