Speech signal endpoint detection method based on dynamic cumulant estimation

Abstract

The invention discloses a speech signal endpoint detection method based on dynamic cumulant estimation, including sliding window-based high-order cumulant recursive computation and sliding window kurtosis-based endpoint detection. The sliding window-based high-order cumulant recursive computation includes: adding a rectangular window to original sample data, subjecting intra-window data to cumulant estimation, updating the intra-window data each time when one sample point is slid, and thus dynamically estimating the cumulant. The sliding window kurtosis-based endpoint detection includes: computing sliding window kurtosis and energy features by the use with the sliding window-based high-order cumulant recursive computation, thus subjecting speech signals to endpoint detection. Compared with the prior art, the method has the advantages that endpoint detection is based on sliding window kurtosis and energy double-threshold, the sliding window kurtosis is highly sensitive to the start point of a speech segment, resistance to noise interference is better, and the method is of better robustness in the noise environment.

Description

technical field [0001] The invention relates to the fields of statistical analysis of data and signal processing, in particular to a method for detecting endpoints of speech signals based on dynamic cumulant estimation. Background technique [0002] With the development of human-computer interaction technology, speech recognition has become the focus of research in the field of artificial intelligence and pattern recognition. Speech is the most important and convenient way for human beings to transmit information, and it is also one of the most direct ways to realize human-computer interaction. It is of great practical significance to allow machines to accurately recognize voice commands and perform corresponding operations. Related research has broad application prospects in many fields such as medicine, military and industry. As the front-end processing of speech recognition, the goal of speech endpoint detection is to distinguish the voiced segment and the unvoiced segme...

AI Technical Summary

Problems solved by technology

Therefore, using all historical data for statistical analysis not only cannot improve the estimation accuracy, but may also cover up the real statistical characteristics of the data.

In addition, in the process of data collection in a real environment, random large-scale outlier interference will cause large errors in statistical analysis results

Since traditional online algorithms rely on all signal data, errors caused by outliers are highly transitive

Images