Real-time speech endpoint detection method and device

Abstract

The invention relates to the technical field of speech, in particular to a real-time speech endpoint detection method and device. The method comprises the following steps that signal framing and emphasis are carried out; pulse removal processing is carried out; direct current components are removed; the short-time energy and zero-crossing rate of each frame of signal are calculated; windowing processing is carried out; spectrum reduction processing is carried out; spectral entropy is calculated; transformation smooth spectral entropy is calculated; a speech frame and a noise frame are preliminarily judged; the transformation smooth spectral entropy and a threshold are processed; a start frame and end frame in a speech segment are judged. The real-time speech endpoint detection method and device have the advantages that according to the conditions under which a signal is judged and a judged result, thresholds of parameters, such as a spectrum reduction threshold, the transformation smooth spectral entropy, the corresponding short-time energy, corresponding short-time average energy and a spectrum reduction power spectrum are weighted and updated, so that the thresholds are more andmore accurate, and finally the judged speech start frame and end frame are also more and more accurate; the method can efficiently and accurately detect speech in real time.

Description

technical field [0001] The invention relates to the technical field of speech, in particular to a real-time speech endpoint detection method and device. Background technique [0002] Voice Activity Detection (VAD), which detects speech segments in a signal, is also a speech endpoint detection technique. Endpoint detection has always been of great significance in the field of speech signal processing. As the front end of speech recognition, accurate endpoint detection can improve the accuracy of recognition; used in speech enhancement systems, it can perform accurate noise model estimation; in the field of speech coding, it can reduce the average bit rate of coding and reduce power consumption. [0003] At present, endpoint detection can be roughly divided into two categories: model-based detection methods and feature-based detection methods. [0004] The model-based method is based on the statistical analysis of data to establish a model that can better describe the intern...

AI Technical Summary

Problems solved by technology

[0006] 1. Based on a single parameter: a method and system for detecting an isolated word speech endpoint disclosed in CN200710179342, based on average energy, for strong noise, the energy method cannot be distinguished; a method for detecting an endpoint of speech recognition disclosed in CN201110071269, based on The linear predictive coding coefficient has a single judgment parameter and requires background noise and speech templates, which are difficult to detect for changing noise and speech;

[0007] 2. Based on a small number of parameter combinations, but the parameters are not very good at distinguishing noise and speech, or are not well calculated: the endpoint detection method, device and speech recognition system based on sliding windows disclosed in CN200410083807 are mainly based on energy and signal Noise ratio, which belongs to a relatively rough method, the signal-to-noise ratio is difficult to estimate accurately, and strong noise is basically difficult to distinguish; CN200410090802 discloses a speech endpoint detection method applied to a speech recognition system, which judges frame by frame and divides frequency bands into different signal-to-noise ratios sub-bands, and then judge the starting frame according to the harmonic characteristics. Similarly, the signal-to-noise ratio is difficult to estimate accurately, and there are few judgment parameters; the endpoint detection system and its processing method based on the fundamental frequency disclosed in CN201410221983 are based on the fundamental frequency, and The harmonic position of the fundamental frequency is assisted, and it is easy to make mistakes in determining the position of the fundamental frequency. For some spectral components and their rich strong noise, such as strong white noise, they cannot be distinguished, and there are still relatively few judgment parameters;

[0008] 3. Based on parameters such as information entropy and spectral entropy, which are easy to calculate and can well distinguish noise and speech, and combined with other parameters, there are many parameters: an adaptive endpoint detection using short-term time-frequency values ​​disclosed in CN201410292519 The method is based on short-term energy, short-term information entropy and relative value of short-term amplitude. Information entropy is processed frequency-by-frequency, and the error is large, and the amount of calculation is large; CN201710086400 discloses a low signal-to-noise ratio environment based on spectral entropy improvement The voice endpoint detection method is mainly based on the subband spectral entropy and energy ratio, but the subbands are not adaptively divided, and no abnormal subband processing is added, and no smoothing and other processing are added.

Images