Abstract

We present an analysis of several publicly available automatic speech recognizers (ASRs) in terms of their suitability for use in different types of dialogue systems. We focus in particular on cloud based ASRs that recently have become available to the community. We include features of ASR systems and desiderata and requirements for different dialogue systems, taking into account the dialogue genre, type of user, and other features. We then present speech recognition results for six different dialogue systems. The most interesting result is that different ASR systems perform best on the data sets. We also show that there is an improvement over a previous generation of recognizers on some of these data sets. We also investigate language understanding (NLU) on the ASR output, and explore the relationship between ASR and NLU performance. 1 Introduction
Dialogue system developers who are not also speech recognition experts are in a better position than ever before in terms of the ease of integrating existing speech recognizers in their systems. While there have been commercial solutions and toolkits for a number of years, there were a number of problems in getting these systems to work. For example, early toolkits relied on specific machine hardware, software, and firmware to function properly, often had a difficult installation process, and moreover often didn’t work well for complex dialogue domains, or challenging acoustic environments. Fortunately the situation has greatly improved in recent years. Now there are a number of easy to use solutions, including open-source systems (like PocketSphinx), as well as cloud-based approaches.

Date

2013

Authors

Fabrizio Morbini, Kartik Audhkhasi, Kenji Sagae, Ron Artstein, Doğan Can, Panayiotis Georgiou, Shrikanth Narayanan, Anton Leuski, David Traum

Conference

Proceedings of the SIGDIAL 2013 Conference

Pages

394-403