Intonation and dialog context as constraints for speech recognition.

Area of Science:

• Computational Linguistics
• Speech Processing
• Artificial Intelligence

Background:

• Automatic Speech Recognition (ASR) systems typically use bigram language models.
• Task-oriented dialog speech presents unique challenges for ASR performance.
• Dialog context and intonation are often underutilized in conventional ASR systems.

Purpose of the Study:

• To improve automatic speech recognition (ASR) performance by incorporating intonation and dialog context.
• To develop and evaluate move-specific language models for task-oriented dialog.
• To investigate the impact of automatic dialog move type recognition on ASR word error rates.

Main Methods:

• Utilized the DCIEM Maptask corpus of spontaneous task-oriented dialog speech.
• Developed separate bigram language models for each of the 12 dialog move types.
• Combined an intonation model, a dialog model, and speech recognizer likelihoods for move type determination.

Main Results:

• Move-specific language models significantly reduced word error rate when the correct model was applied.
• The integrated system, combining automatic move type recognition with move-specific models, reduced overall word error rate compared to a baseline.
• Word error rate improvements were primarily observed in 'initiating' move types, while 'response' move types showed good recognition of the type itself.

Conclusions:

• Integrating intonation and dialog context, through move-specific language models and dialog act recognition, offers a significant improvement for ASR in task-oriented dialog.
• The approach effectively enhances word recognition for critical dialog segments.
• Further research can explore more sophisticated dialog and intonation modeling for broader ASR applications.