Recognition of spontaneous speech

• Authors: Peter Stubley
• Venue: Transcribe '98 Proceedings of the Workshop on Partially Automated Techniques for Transcribing Naturally Occurring Continuous Speech
• Year: 1998

Current speech recognition systems are capable of performing complex tasks for co-operative users by determining their requirements through a conversation. Most systems have been constructed without attempting to accurately model spontaneous speech. Some components, such as the parser, can be easily made robust to some of the artifacts of conversational speech. Others, such as the pronunciation models, simply ignore the possibility that incomplete words can occur. This results in some recognition errors, and may cause the application to begin to perform the wrong the action. Typically, however, the next several conversation turns can identify and correct the error. This talk gives a brief overview of state-of-the-art of spoken language systems and describes how some of the components are affected by artifacts of spontaneous speech. Large bodies of accurately transcribed spontaneous speech are required to learn the properties of spontaneous events.