语音端点检测方法研究

发布时间：2018-01-10 02:30

本文关键词：语音端点检测方法研究　出处：《大连理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：语音端点检测是指在背景噪声存在下,找到一个单词或语音段的起始点和结束点。它在语音识别、语音编码和语音传输等语音信号处理中起着重要作用。一种有效的端点检测方法,不仅能正确标识语音端点,而且能减少数据处理时间、提高效率和节省数据存储空间。本文介绍了语音信号的产生模型和人耳听觉特性,分析了语音分帧和加窗这些预处理过程。简要回顾了常用的端点检测方法,如短时能量和过零率法、谱方差法、MFCC倒谱距离法和谱熵法等。总结了各种常用算法的思想、流程及特点,并给出了各算法所用特征参数曲线图。提出了两种新的端点检测算法：(1)提出了一种改进的自适应子带谱方差端点检测方法。因为不同类型噪声的频率能量集中在不同的频率带上,所以可以通过准确地去掉这些含有太多噪声的频率带,并自适应地选择有用子带能量,来得到更多准确的语音频率谱信息。使语音段方差和噪声段方差之间的区分度变大,端点检测更容易,准确率更高。同时有效减少了需处理的语音数据,提高系统整体性能。并将该算法应用到高斯连续HMM语音识别系统中,实验表明,该算法提高了识别系统准确率及抗噪性。(2)提出了一种无前导静音段语音的端点检测算法。基于阈值的端点检测方法假设语音起始段为无语音的静音段,并根据该无语音的起始段的特征值设置阈值。如果检测的语音帧的特征值大于阈值,就认为该帧是语音帧,否则认为该帧是噪声帧。若该假设不成立,预设的阈值将不可用,最终导致端点检测错误。本文通过FCM聚类方法避开了阈值设置问题。该算法在检测无前导静音段语音时,表现出了很好的性能。
[Abstract]:Speech endpoint detection refers to finding the beginning and ending point of a word or speech segment in the presence of background noise. Speech coding and voice transmission play an important role in speech signal processing. An effective endpoint detection method can not only correctly identify speech endpoints, but also reduce data processing time. This paper introduces the speech signal generation model and auditory characteristics, analyzes the preprocessing process of speech framing and windowing, and briefly reviews the common endpoint detection methods. Such as short-time energy and zero-crossing rate method, spectral variance method, MFCC cepstrum distance method and spectral entropy method, etc. The ideas, flow chart and characteristics of various commonly used algorithms are summarized. The graph of characteristic parameters used in each algorithm is given, and two new endpoint detection algorithms: 1) are proposed. An improved adaptive subband spectrum variance endpoint detection method is proposed because the frequency energy of different types of noise is concentrated on different frequency bands. So it is possible to remove the frequency bands which contain too much noise accurately and to select the useful sub-band energy adaptively. In order to get more accurate speech frequency spectrum information, the distinction between the variance of speech segment and the variance of noise segment becomes larger, the endpoint detection is easier and the accuracy is higher. At the same time, the speech data that needs to be processed is reduced effectively. The algorithm is applied to Gao Si continuous HMM speech recognition system. This algorithm improves the accuracy and anti-noiseness of the recognition system. (2) an endpoint detection algorithm for speech without leading mute segment is proposed. The threshold based endpoint detection method assumes that the beginning segment of speech is a voiceless mute segment. If the eigenvalue of the detected speech frame is greater than the threshold value, the frame is considered to be a speech frame, otherwise the frame is considered to be a noisy frame. The preset threshold will not be available, resulting in endpoint detection errors. In this paper, the threshold setting problem is avoided by FCM clustering method. The algorithm shows good performance in detecting speech without leading mute segment.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.3

【相似文献】