A New Acoustic-Based Pronunciation Distance Measure

Area of Science:

• Linguistics
• Speech Science
• Acoustic Phonetics

Background:

• Assessing foreign accent strength is crucial for language learning and communication.
• Current methods rely on phonetic transcription, which is labor-intensive and prone to errors.
• There is a need for efficient and objective measures of accent strength.

Purpose of the Study:

• To develop and validate an acoustic-only distance measure for quantifying foreign accent strength in American-English.
• To compare the performance of the acoustic measure against human judgments and existing transcription-based methods.
• To investigate the sensitivity of the acoustic measure to various phonetic and prosodic features.

Main Methods:

• Developed an acoustic distance measure using dynamic time warping (DTW) on speech features.
• Applied speaker-based cepstral mean and variance normalization to minimize speaker variability.
• Collected speech samples from non-native and native American-English speakers.
• Correlated acoustic distances with native speaker ratings of accent strength (n > 1,100).

Main Results:

• The acoustic measure showed a strong negative correlation (r = -0.71, p < 0.0001) with human judgments of native-likeness.
• Performance was comparable to state-of-the-art transcription-based methods (r = -0.77).
• The measure demonstrated sensitivity to segmental, intonational, and durational differences, but was affected by recording device variations.

Conclusions:

• The proposed acoustic distance measure provides a convenient and effective tool for assessing foreign accent strength in American-English.
• It offers a viable alternative to time-consuming phonetic transcription methods.
• Further research may refine the measure to mitigate variations caused by recording equipment.