UNCONSTRAINED DYSFLUENCY MODELING FOR DYSFLUENT SPEECH TRANSCRIPTION AND DETECTION.

Area of Science:

• Speech processing
• Computational linguistics
• Phonetics

Background:

• Accurate transcription and detection of dysfluent speech are crucial for speech modeling.
• Current methods often focus on either transcription or detection, with limited performance.
• Extensive manual annotation is a bottleneck in dysfluency research.

Purpose of the Study:

• To present an unconstrained dysfluency modeling (UDM) approach for automatic and hierarchical dysfluency transcription and detection.
• To reduce the reliance on manual annotation in dysfluency modeling.
• To improve phonetic transcription capabilities for dysfluent speech.

Main Methods:

• Developed an unconstrained dysfluency modeling (UDM) approach.
• Implemented a hierarchical method for simultaneous transcription and detection.
• Introduced a simulated dysfluent dataset, VCTK++, for phonetic transcription enhancement.

Main Results:

• The UDM approach effectively addresses both transcription and detection tasks.
• Experimental results show the method's effectiveness and robustness.
• The VCTK++ dataset aids in improving phonetic transcription accuracy.

Conclusions:

• The proposed UDM approach offers a comprehensive solution for dysfluent speech modeling.
• The method demonstrates significant improvements in both transcription and detection.
• UDM reduces the need for manual annotation, facilitating further research.