【作者】 郝静；

【导师】 张刚；

【作者基本信息】 太原理工大学 ， 信号与信息处理， 2008， 硕士

【摘要】 多参数结合的实时语音分段算法在语音信号处理领域中具有十分重要的意义。本文的主要研究目的是寻找一种有效的方法,能够将多种语音信号的特征参数结合起来,实时地检测或确定连续语音信号采样的突变点,并以此突变点对语音信号分段。本文首先对几种常用于语音分段算法的特征参数进行了分析,提出了实时改进算法。采用帧长20样点,利用140个历史样点构成长度为160点的窗进行处理。并在此基础上提出了一种实时自相关的语音分段算法,实验表明,改进后的方法能够实现对语音信号的实时处理,提取的特征参数能够表征语音信号的突变信息,对不同信噪比的语音信号进行分段,其准确率能达到80%以上。本文在实现了特征参数实时提取的基础上,研究分析了粒计算理论,将粒理论应用到语音分段技术上,提出了一种基于粒计算的语音分段算法。该算法利用粒计算处理不精确、模糊、不确定、部分真实和海量的信息的能力对8种特征参数进行了分析,得到它们之间的相互关系及对语音分段的重要度,最终构成决策规则,从而有效的进行语音分段判决。实验发现该算法能够将多种特征参数的优点结合起来,利用其决策规则可以找到大部分语音信号的分段点,但存在错判漏判情况,判决准确度不高。针对错判和漏判的问题,经过大量的实验分析,本文对基于粒计算的语音分段算法提出了改进方案。改进从两个方面进行:一是对特征参数采集过程的改进,目的是消除噪声干扰,增加决策规则的准确性;二是对判决过程的改进,通过在决策规则的基础上加入自相关与能量参数构成双路径判决规则,达到辅助判决减少漏判错判的目的。改进后的算法对不同信噪比的语音数据的判决准确率均达到90%以上,漏判率和错判率明显降低,是一种比较理想的多参数结合语音实时分段算法。更多还原

【Abstract】 Real-Time Multi-Parameter Speech Segmentation Algorithm has very important meaning in the field of Speech Signal Processing. This research tries to find an effective method combines various characteristic parameters of speech signal to detect the sharp variations in continuous speech signal. And we can carry out the segmentation by the sharp variations of speech signal.First, in analyzing several characteristic parameters in common use, this paper presents a real time improvement algorithm, which processes in the window of 160 sample points including current 20 sample points frame and history 140 sample points, and based on the algorithm, this paper also present the Real-time Autocorrelation Speech Signal Segmentation Algorithm. Research illustrates the real time improved algorithm can process the real time speech signal. The characteristic parameters can characterize the sharp variations of speech signal. The real time improvement algorithm can segment various speech signal of signal noise ratio; the precision ratio of decisions reaches at 80 percent.Based on the realization of characteristic parameters distilling and analyzing of the granular computing theory, this paper applies it on the theology of speech signal segmentation and presents the Speech Signal Segmentation Algorithm Based on Granular Computing。This algorithm takes the advantage of the abilities on processing inaccurate, fuzzy, undefined, partial real and mass information to analyze eight characteristic parameters of speech signal, then gets the relationship between the characteristic parameters and the significance of the characteristic parameters, and finally forms the decision rules to segment speech signal efficiently. Research finds this algorithm can combine many kinds of characteristic parameters’ benefits on speech segmentation. It can find almost all the segmentation points of speech signal by its generated decision rules; however, some missed and inaccuracy decisions also exist.To resolve the missed and inaccuracy arbitration problem, through a great deal of experimentations, this paper present the two improve methods of the Speech Signal Segmentation Algorithm Based on Granular Computing: First, improves the characteristic sampling process in order to eliminate noise disturb and increase the precision of arbitration; Second, on the foundation of arbitration rule, adds autocorrelation and energy parameters and forms two-way arbitration rule, this improvement can assist the arbitration and reduce the missed decisions. After these improvements, the precision of arbitration reaches the 90 percent, the missed and inaccuracy decisions obviously decrease. In conclusion, the Speech Signal Segmentation Algorithm Based on Granular Computing is a relative perfect Real-Time Multi-Parameter Speech Segmentation Algorithm.更多还原