Automatic measurement of voice onset time using discriminative structured prediction

Area of Science:

• Speech processing
• Acoustic phonetics
• Machine learning

Background:

• Voice onset time (VOT) is a crucial phonetic feature for distinguishing speech sounds.
• Manual measurement of VOT is time-consuming and subject to inter-annotator variability.
• Automatic methods for VOT measurement are needed for large-scale speech analysis.

Purpose of the Study:

• To develop a discriminative large-margin algorithm for automatic VOT measurement.
• To predict VOT as a structured output from speech segments.
• To achieve performance comparable to human annotators.

Main Methods:

• A large-margin algorithm was trained using manually labeled speech data.
• The algorithm predicts VOT from acoustic features derived from spectral and temporal cues.
• Feature functions mimic those used by human VOT annotators.

Main Results:

• The algorithm achieved performance near human inter-transcriber reliability across four speech corpora.
• The method demonstrated favorable comparison with previous automatic VOT measurement techniques.
• Performance remained consistent when training and testing on different corpora.

Conclusions:

• The developed algorithm offers a practical and accurate solution for automatic VOT measurement.
• The method shows robustness to variations in speech corpora and requires minimal training data (50-250 examples).
• This approach has significant potential for applications in speech research and technology.