SBIR Phase I: Tools for Information Retrieval and Document Classification Using Fast Phonetic Word-Spotting Technology

这个小型企业创新研究第一阶段的研究项目将进行必要的研究和开发，以大大提高快速语音单词识别器的信息检索能力。完成的研究将为低质量电话音频或多媒体数字源的语音文档检索和分类提供新的方法。语音文档检索一直是广播新闻领域研究的热点问题。然而，在许多应用程序中，用户必须检索和分类质量较低的音频文档。最常用的方法包括将音频流或文件转换为假设的单词序列（语音到文本或STT），然后使用基于文本的信息检索。尽管这已被证明对广播新闻文档检索是有效的，但它也有缺点。例如，STT对语言模型的明确使用将假设的单词序列限制为其词典中的那些。另一方面，语音匹配能够识别词典中不存在的关键字的可能实例，例如名称。STT方法的一个优点是基于文本的信息检索方法的适用性，这种方法在错误率相当小的高质量音频上工作得很好。然而，在高容量的电话信道中，更好的解决方案是必要的，因为计算负担和低精度使得STT不切实际。该计划的目标是研究和开发基于语音的文档检索和分类算法。在现有的大型语料库上比较基于语音搜索的检索系统的适用性。所提出的研究的关键创新是使搜索技术适应音频存在而文本不存在的环境。科学地说，算法必须在概率框架中工作，因为语音单词识别总是基于置信度度量。商业上，现有的多媒体或音频档案将可用于数据挖掘。此外，文档类型的决定（例如，呼叫中心的电话是否为投诉？）在市场情报、安全分析、质量分析和任何呼叫隔离应用程序中开放商业应用程序。

This Small Business Innovation Research Phase I research project will perform the research and development necessary to greatly enhance the information retrieval capability of a fast phonetic word-spotter. The completed research will lead to new methods for spoken document retrieval and classification on low quality telephony audio or multimedia digital sources. Spoken document retrieval has been a well-researched problem in the domain of broadcast news. However, many applications exist where users must retrieve and classify documents with lower quality audio. The most commonly applied method involves converting an audio stream or file into a hypothesized sequence of words (Speech-to-Text or STT), and subsequently using text- based information retrieval. Although this has been shown to be effective for broadcast news document retrieval, this has drawbacks. For example, STT's explicit use of language models limits the hypothesized word sequences to those within its lexicon. On the other hand, phonetic matching is capable of identifying likely instances of keywords, such as names, which are not in a lexicon. One advantage of the STT approach is the applicability of text-based information retrieval methods, which work well on high quality audio where the error rates are fairly small. However, better solutions are necessary over a high volume telephony channel where the computational burden and low accuracy make STT impractical. The goal of the proposed project is to research and develop phonetic-based document retrieval and classification algorithms. The applicability of retrieval systems based on phonetic searches will be compared on large existing corpora.The key innovation of the proposed research is to adapt search techniques to function in environments where audio exists, but text does not. Scientifically, algorithms must be made to work in a probabilistic framework, since phonetic word spotting is always based on confidence measures. Commercially, existing multimedia or audio archives will be available for data mining. In addition, decisions of document type (e.g., was the phone call to the call center a complaint?) open commercial applications in market intelligence, security analysis, quality analysis, and any call segregation application.