Speech technology today has a wide variety of existing and potential applications in so many areas of our life. From dictating systems to voice translation, from digital assistants like Siri, Google Now, and Cortana, to telephone dialogue systems. Many of these applications have to rely on an Automatic Speech Recognition (ASR) component. This component not only has to perform well, but it also has to perform well in adverse environments. After all, a dictating system which requires that we insulate our office, or a digital assistant that cannot work in traffic, or in a room full of chatting people is not so helpful. For this reason, noise robust ASR has been a topic of intensive research. Yet, human-equivalent performance has not been achieved. This motivated many to search for ways to improve the robustness of automatic speech recognition based on human speech perception. One popular method inspired by the examination of the receptive fields of auditory neurons is that of spectro-temporal processing.ududIn spectro-temporal processing, the aim is to capture the spectral and temporal modulations of the signal simultaneously. One simple way to do so is to extract the features to be used from spectro-temporal patches, and then use the resulting features in the same manner one would use traditional features like MFCCs. There is more than one way to bake a cake, however. And in this case this is true twice over. For one, there are various ways to extract our features from the patches. But there are other, more sophisticated ways to incorporate the concept of spectro-temporal processing into a speech recognition system. In this study we examine many such methods -- some simpler, some more sophisticated, but all stemming from the same basic idea. By the end of this study we will demonstrate that these methods can indeed lead to more robust speech recognition. So much so, that they can provide results that are competitive with the state-of-the-art results.
展开▼
机译：当今的语音技术在我们生活的许多领域中具有广泛的现有和潜在应用。从口述系统到语音翻译，从Siri，Google Now和Cortana等数字助理到电话对话系统。这些应用程序中的许多都必须依赖于自动语音识别（ASR）组件。该组件不仅必须具有良好的性能，而且还必须在不利的环境中具有良好的性能。毕竟，要求我们隔离办公室，不能在交通中不能工作的数字助理或在闲聊的房间中的命令系统并没有太大帮助。因此，抗噪ASR一直是深入研究的主题。但是，尚未实现与人类相当的性能。这激发了许多人寻求基于人类语音感知来提高自动语音识别的鲁棒性的方法。一种受听觉神经元接受域检查启发的流行方法是光谱时间处理。 ud ud在光谱时间处理中，目的是同时捕获信号的频谱和时间调制。一种简单的方法是从频谱时态补丁中提取要使用的功能，然后以与使用诸如MFCC之类的传统功能相同的方式使用生成的功能。但是，烘烤蛋糕有多种方法。在这种情况下，这是正确的两次。首先，有多种方法可以从补丁中提取我们的功能。但是，还有其他更复杂的方法可以将频谱时处理的概念整合到语音识别系统中。在本研究中，我们研究了许多这样的方法-一些更简单，更复杂的方法，但是所有这些方法都是基于相同的基本思想。到本研究结束时，我们将证明这些方法确实可以导致更强大的语音识别。如此之多，他们可以提供与最新结果相抗衡的结果。
展开▼
