语音可懂度客观评价策略的研究

发布时间：2018-07-13 14:03

【摘要】：在信息时代背景下,语音是人们日常生活中最直接的信息传递与交换形式。而现实生活中,多数语音往往伴随着噪声,使人感到舒适度下降。听力受损的人理解不同背景噪声下的语音是很困难的。因此,怎样从信号中分离噪声、改善可懂度以及如何评价可懂度成为了一个重要问题。目前,理想二值掩蔽(Ideal Binary Mask, IBM)技术的提出,给噪声背景下评价语音可懂度提供了希望。最近研究表明每个时间-频率单元对语音可懂度重要性与语音内容相关。时间-频率单元被分为噪声主导和目标语音主导两类。本文实验结果表明每个语音存在的时间-频率单元对语音可懂度的重要性与目标语音的响度有关。2008年,Li和Loizou研究证明,当输入信噪比在-5dB下,误报错误比丢失错误对语音可懂度的影响更大。本实验对输入混合信噪比进行了扩充,研究了不同输入混合信噪比下,两种掩蔽错误对语音可懂度的影响。本文实验证明,当混合信噪比在0dB以下时,误报错误比丢失错误对语音可懂度影响更大。理想二值掩蔽的研究大都是基于这样一个假设：每个语音分离单元对语音整体可懂度的贡献度是一样的,但没有改变IBM结构。考虑到两种掩蔽错误对语音可懂度影响程度和IBM结构对语音可懂度的影响,本文提出了对丢失错误和误报错误进行加权控制的方法,同时结合IBM结构的改变,意在提出更好的评价语音可懂度的指标。
[Abstract]:In the background of the information age, speech is the most direct form of information transfer and exchange in people's daily life. In real life, most of the voice often accompanied by noise, making people feel comfortable. It is difficult to understand the sound of the hearing impaired. For this, how to separate the noise from the signal and improve the intelligibility. And how to evaluate intelligibility has become an important problem. At present, the proposed Ideal Binary Mask (IBM) technology provides hope for the evaluation of speech intelligibility under the noise background. Recent studies have shown that the importance of each time frequency unit to the speech intelligibility is related to the speech content. The time frequency unit is divided into noise, and the time frequency unit is divided into noise. The experimental results show that the importance of the time frequency unit on speech intelligibility in each speech is related to the phonetic loudness of the target speech in.2008 years. Li and Loizou studies show that when the input signal-to-noise ratio is in -5dB, the error of false alarm is greater than that of the loss and error on the speech intelligibility. The mixed SNR is expanded to study the effect of two masking errors on speech intelligibility under different input mixed signal to noise ratio. In this paper, it is proved that when the mixed SNR is below 0dB, false alarm error has greater influence on speech intelligibility than loss error. The study of ideal two value masking is based on such a hypothesis. The contribution of a speech separation unit to the overall intelligibility of speech is the same, but it does not change the IBM structure. Considering the effect of the two masking errors on the degree of speech intelligibility and the effect of the IBM structure on the intelligibility of speech, this paper proposes a weighted control method for the loss and misrepresentation errors, and combines the changes of the IBM structure. A better indicator to evaluate the intelligibility of speech.
【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.3

【相似文献】