Automatic measurement of vowel duration via structured prediction

Area of Science:

• Linguistics
• Computational Linguistics
• Speech Science

Background:

• Phonetic studies often rely on manual annotation, which is subjective, time-consuming, and limits scalability.
• Accurate and automated measurement of phonetic features like vowel duration is crucial for replicable research.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for the automatic measurement of vowel duration.
• To provide a scalable and replicable method for phonetic analysis, reducing reliance on manual annotation.

Main Methods:

• A machine learning model based on the structured prediction framework was trained using manually-annotated vowel duration data.
• The model processes acoustic signal segments, extracting acoustic features to map inputs into a vector space.
• The algorithm was trained to minimize discrepancies between predicted and manually measured vowel durations.

Main Results:

• The developed model automatically estimates vowel duration without requiring phonetic or orthographic transcriptions.
• Comparison with manual annotations indicated the model outperformed a hidden Markov model-based forced aligner, the previous gold standard.

Conclusions:

• The machine learning algorithm offers an effective and automated solution for vowel duration measurement.
• This approach enhances the scalability and replicability of phonetic studies by removing the need for manual annotation.