Building searchable collections of enterprise speech data

Abstract

We have applied speech recognition and text-mining technologies to a set of recorded outbound marketing calls and analyzed the results. Since speaker-independent speech recognition technology results in a significantly lower recognition rate than that found when the recognizer is trained for a particular speaker, we applied a number of post-processing algorithms to the output of the recognizer to render it suitable for the Textract text mining system. We indexed the call transcripts using a search engine and used Textract and associated Java technologies to place the relevant terms for each document in a relational database. Following a search query, we generated a thumbnail display of the results of each call with the salient terms highlighted. We illustrate these results and discuss their utility. We took the results of these experiments and continued this analysis on a set of talks and presentations. We describe a distinct document genre based on the note-taking concept of document content, and propose a significant new method for measuring speech recognition accuracy. This procedure is generally relevant to the problem of capturing meetings and talks and providing a searchable index of these presentations on the web.