Automatic detection of prosodic boundaries in spontaneous speech

Area of Science:

• Computational Linguistics
• Speech Processing
• Human-Computer Interaction

Background:

• Automatic speech recognition (ASR) and natural language processing (NLP) require effective parsing of conversational speech.
• Identifying prosodic phrase boundaries is essential for accurate speech parsing but remains challenging for automated systems.
• Existing methods often rely on complex linguistic and acoustic cues, necessitating model training.

Purpose of the Study:

• To develop a simple, reliable, and training-free method for automatically detecting prosodic phrase boundaries in conversational speech.
• To leverage readily available ASR output for prosodic boundary identification.
• To evaluate the effectiveness of the proposed method in producing syntactically valid and acoustically coherent phrases.

Main Methods:

• Utilized two prosodic cues derived from ASR output: speech rate discontinuities (pre-boundary lengthening, phrase-initial acceleration) and silent pauses.
• Developed a method that does not require prior model training.
• Identified phrase boundaries based on detected discontinuities and pauses.

Main Results:

• The proposed method successfully identified prosodic phrase boundaries.
• The resulting phrases demonstrated syntactic validity and pitch reset.
• The automated boundary detection compared favorably with manual tagging by human annotators.
• The method supports the concept of prosodic phrases as coherent textual and acoustic units.

Conclusions:

• The developed method offers an effective, simple, and reliable approach to automatic prosodic phrase boundary detection.
• This technique can significantly benefit ASR and NLP applications by improving conversational speech parsing.
• The findings reinforce the linguistic and acoustic coherence of prosodic phrases.