语音增强技术研究及在语音识别中的应用

发布时间：2019-04-16 14:43

【摘要】：语音是生活中最重要的信息载体,然而它很容易受到噪声的污染,从而干扰信息的传递。语音增强,它能将重要的语音信号从嘈杂的带噪语音中分离出来,将噪声的影响尽可能地减少。语音识别技术在现实生活中应用越来越广泛,安静环境中的语音识别技术已经日趋成熟,而噪声环境下的语音识别技术仍然是一个难题。语音识别系统要实现抗噪就必须要尽量减少噪声的干扰,因此将语音增强技术结合到语音识别中来实现识别系统的抗噪性,具有很大的研究意义。本文的研究工作主要分成三个部分。第一,深入研究了几种传统的语音增强算法,包括谱减法、维纳滤波和最小均方误差估计算法,并在此研究基础上对这些算法进行了详细的仿真实验。基于实验的基础上,对比分析这些算法各自的优缺点,并从信噪比和语音失真两方面去评价这些语音增强方法。最后,对各种算法加以总结得出不同噪声要求的情况下比较合适的语音增强算法。第二,对语音识别技术进行了分析和比较研究,着重分析了预处理、端点检测等几个语音识别实现需要解决的关键性技术。在此基础上,我们实现了主流的动态时间规整和隐马尔可夫算法。从识别率上,对两种算法进行了分析,并实验比较了这些语音识别系统抗噪性的差异。第三,提出了一种基于语音增强技术的语音识别算法。该算法将语音增强与语音识别进行联合处理,即在识别之前先进行语音增强处理。实验证明,实现的语音识别系统具有良好的抗噪性。通过比较几种语音增强算法在语音识别上应用结果及分析,得出了语音识别应用中较好的语音增强算法。设计了一个多引擎的语音识别系统,该系统抗噪性和识别率上,都比传统的识别算法效果更出色。
[Abstract]:Speech is the most important information carrier in life. However, it is easy to be polluted by noise, which interferes with the transmission of information. Speech enhancement, which can separate important speech signals from noisy speech, and reduce the influence of noise as much as possible. Speech recognition technology is more and more widely used in real life. Speech recognition technology in quiet environment has become more and more mature, and speech recognition technology in noise environment is still a difficult problem. If speech recognition system is to resist noise, it is necessary to reduce noise interference as far as possible. Therefore, it is of great significance to combine speech enhancement technology with speech recognition to realize the anti-noise property of speech recognition system. The research work of this paper is divided into three parts. Firstly, several traditional speech enhancement algorithms, including spectral subtraction, Wiener filtering and least mean square error estimation (MMSE), are deeply studied, and detailed simulation experiments are carried out on the basis of this study. On the basis of experiments, the advantages and disadvantages of these algorithms are compared and analyzed, and the speech enhancement methods are evaluated from two aspects: signal-to-noise ratio (SNR) and speech distortion. Finally, various algorithms are summarized to get a more suitable speech enhancement algorithm under different noise requirements. Secondly, the speech recognition technology is analyzed and compared, and the key technologies to realize speech recognition, such as preprocessing and endpoint detection, are emphatically analyzed. On this basis, we implement the mainstream dynamic time regularization and hidden Markov algorithm. In terms of recognition rate, the two algorithms are analyzed, and the differences of anti-noise performance of these speech recognition systems are compared experimentally. Thirdly, a speech recognition algorithm based on speech enhancement technology is proposed. The algorithm combines speech enhancement with speech recognition, that is, speech enhancement before recognition. The experimental results show that the proposed speech recognition system has good anti-noise property. By comparing the application results and analysis of several speech enhancement algorithms in speech recognition, a better speech enhancement algorithm in speech recognition application is obtained. A multi-engine speech recognition system is designed. The anti-noise performance and recognition rate of the system are better than the traditional recognition algorithm.
【学位授予单位】：福州大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【相似文献】