Feature-Combination for Noise Robust Speech Pattern Processing

当前的计算机语音自动识别系统在精心控制的环境下也能获得令人满意的性能。然而，在现实生活中，语音信号通常会受到背景环境噪声的污染。虽然人类对噪声表现出很强的鲁棒性，但目前的自动语音识别系统的性能下降很快，即使是数字识别这样的简单任务。语音信号可以由多个特征表示，这些特征可以通过在特定的信息源上使用不同的信息源或不同的处理技术来获得。在给定的一组特征中，可能存在一些被噪声破坏的特征。理想情况下，受噪声影响的特征应该排除在识别之外。为了实现这一点，需要先验地了解噪声特征的同一性。不幸的是，如果没有关于噪声的先验信息，定位损坏的特征本身可能是一项困难的任务。因此，为了挖掘未受影响的特征的潜力，我们面临的问题是如何在不知道噪声的情况下组合特征。在我们之前的工作中，我们开发了一个特征组合模型，试图释放对噪声特征识别的需求。先前研究的一个关键结果是，当噪声具有部分频率/时间特征时，不使用噪声特征信息的模型与使用关于噪声特征的全部先验知识的模型取得了相似的识别性能。我们以前的研究是在不知道噪声的情况下，为了消除噪声特征的影响，处理了一个一般的特征组合问题。这为开发更强大的特征组合模型提供了良好的基础，这些模型能够利用语音信号的固有属性。我们的研究目标是建立特征组合模型，该模型考虑：(1)在宽带噪声环境下，频谱的谷值容易被噪声破坏，而峰值通常受噪声影响较小；(2)关于特征可靠性的任何信息，这些信息通常可以通过利用语音信号的特性来获得。此外，基于分别对滤波器和源信息建模的语音信号建模研究可以纳入特征组合模型。这些模型将为语音模式处理量身定制，因此应该提供更好的识别性能。我们的最终目标是展示语音和说话人识别方面的竞争力；我们的目标是在标准数据集（分别为TIDIGITS、TIMIT、Resource Management和Switchboard）上实现显著的性能改进。

Current systems for automatic speech recognition by computer can obtain an acceptable performance in carefully controlled environments. However, in real-world situations, speech signal is usually contaminated by an acoustic background environmental noise. While humans show strong robustness to noise, the performance of current automatic speech recognition systems degrades rapidly, even for a simple task such as digit recognition.Speech signal may be represented by multiple features, which may be obtained by using different sources of information or different processing techniques on a specific source. In a given set of features, there may be some features corrupted by noise. Ideally, the features dominated by noise should be excluded from recognition. To achieve this, a-priori knowledge about the identity of the noisy features is required. Unfortunately locating the corrupted features itself can be a difficult task, if there is no prior information about the noise. Thus, to exploit the potential of the unaffected features, we face the problem of how to combine the features when assuming no knowledge about the noise.In our previous work, we developed a feature-combination model that attempts to release the need for identification of the noisy features. A key result of previous studies is that, when the noise has a partial frequency/temporal character, this model using no information about noisy features has achieved similar recognition performance as a model using full a-priori knowledge about the noisy features.Our previous study dealt with a general problem of combination of features in order to eliminate the effect of noisy features under the assumption of no knowledge about the noise. This provides a good base for the development of more powerful feature-combination models capable of exploiting the inherent properties of speech signals. Our proposed research aims to develop feature-combination models that incorporate: (1) the fact that in a wide-band noisy environment, the valleys of spectrum are easily corrupted by noise while peaks are often affected little; (2) any information about reliability of features, which may often be available by exploiting properties of speech signals. Moreover, the proposed investigation on modelling of speech signals based on modelling the filter and source information separately can be incorporated into the feature-combination models. Such models will be tailored for speech pattern processing and thus should provide an improved recognition performance. Our final goal is to demonstrate competitive performance in speech and speaker recognition; we aim to achieve significant performance improvements on standard datasets (TIDIGITS, TIMIT, Resource Management, and Switchboard, respectively).