TL;DR
This paper introduces a novel method for classifying English nativeness using suprasegmental features extracted at the "Accent Group" (AG) level, outperforming traditional acoustic feature-based methods. By focusing on prosodic elements like intonation and rhythm within these natural speech units, the approach achieves high accuracy with significantly fewer features, suggesting the critical role of suprasegmentals in perceived nativeness.
What Problem Does This Paper Address?
Learning to master the suprasegmental aspects of a new language – rhythm, stress, and intonation – is notoriously difficult for learners. These prosodic features are crucial for intelligibility and perception, yet traditional language learning tools often struggle to provide nuanced feedback on them. Existing automatic assessment methods either rely on direct comparisons to specific native speaker examples (limiting generalizability) or use low-level acoustic features that may not fully capture the complex prosodic patterns learners struggle with. This paper aims to bridge this gap by developing a robust and generalizable method for classifying speaker nativeness based on these higher-level prosodic units.
Methods
The core innovation lies in analyzing suprasegmental features at the Accent Group (AG) level. An AG is defined as an accented syllable followed by all unaccented syllables until the next accent or phrase boundary. This unit is chosen because it's large enough to capture prosodic variability but abstract enough for generalizability.
The authors extract three categories of features from each AG:
- Intonation: Modeled using polynomial approximations of the fundamental frequency (f0) contour within the AG. Coefficients from these polynomials form a feature vector.
- Energy: Similar to intonation, energy contours are also modeled with polynomials, generating another feature vector.
- Duration: Relative durations of accented and unaccented syllables within the AG, normalized by the total AG duration, along with the total number of syllables in the AG.
These features are concatenated into an 11-dimensional vector per AG. A Random Forest classifier is then used as a "back end" to combine AG-level scores (derived from GMMs and Decision Trees) into an overall utterance-level nativeness score.
Datasets
The study uses a corpus of parroted speech data, consisting of:
- Japanese Learner Parroting: Recordings from native Japanese speakers repeating English phrases.
- Native English Speaker Parroting: Recordings from native English speakers repeating the same phrases, instructed to match the rhythm and intonation of the original prompts.
The corpus includes statistics on the number of female/male speakers and total phrases/hours for both learner and native groups. The reference prompts were originally from the Rosetta Stone American English product.
Key Insights and Ideas
- Accent Groups as Key Units: The study strongly suggests that the Accent Group is a highly effective unit for analyzing and classifying nativeness, capturing crucial suprasegmental information.
- Prosody Over Acoustics: Suprasegmental features, particularly those related to intonation and rhythm at the AG level, are shown to be highly discriminative for nativeness classification, even outperforming traditional spectral features (PLP) used in a baseline system.
- Dimensionality Reduction: The proposed method uses a remarkably low-dimensional feature set (11 features per AG, with up to 3 AGs per utterance) compared to the baseline's high-dimensional acoustic features (39 PLP features, leading to thousands of dimensions for GSV). This highlights the efficiency and power of prosodic analysis.
- Random Forest Efficacy: The Random Forest classifier proved to be a strong choice for combining AG-level scores, yielding superior results compared to individual GMM or Decision Tree models.
Key Results
The proposed prosodic nativeness classifier achieved:
- 83.3% accuracy for 2-AG utterances.
- 89.1% accuracy for 3-AG utterances.
These results significantly exceed the baseline Gaussian Supervector (GSV) system's performance by more than 10% absolute. The AG-based methods also showed better performance on the AG-level classification task itself compared to the GSV baseline.
Limitations and Future Work
- Accent Group Definition: A limitation acknowledged is that AGs were manually annotated by experts based on prompt acoustics. Future work should focus on automatic AG segmentation for greater flexibility and scalability.
- L1 Diversity: The study primarily used Japanese learners. Future research should explore the proposed method with a wider variety of nonnative English speakers from different L1 backgrounds.
- Feature Set Exploration: Further examination of the relationship between AG features and those derived from pseudosyllables is suggested. Exploring features based on other prosodic units is also a promising direction.
- Application to Unscripted Speech: While the current corpus is based on parroted speech, applying this method to more spontaneous, unscripted learner utterances would be a valuable next step.
Summary updated