TL;DR
This paper introduces Functional Data Analysis (FDA) as a novel, data-driven approach to analyzing intonational contours in speech. By treating pitch contours as continuous functions rather than discrete points, FDA offers a less theory-dependent way to uncover phonetic variations, potentially bridging the gap between phonological theory and acoustic realization.
What Problem Does This Paper Address?
Intonational phonology, particularly in English, has largely relied on autosegmental-metrical theories (like TOBI) that describe pitch contours using discrete high (H) and low (L) tones. While this framework aims for phonetic transparency, the manual labeling of these tones by human annotators is inherently subjective and prone to bias and inter-annotator disagreement. This reliance on hand-labeling can obscure subtle phonetic variations that might be crucial for a complete understanding of intonation. The authors aim to address this by using FDA to analyze entire pitch contours as functions, allowing for a more objective and comprehensive characterization of intonational patterns.
Methods
The core of this paper's methodology lies in the application of Functional Data Analysis (FDA), specifically Functional Principal Component Analysis (FPCA). FDA extends traditional statistical methods to analyze data that are functions or curves. In this context, fundamental frequency (F0) contours are treated as continuous functions.
Here's a breakdown of the process:
- Data Preparation:
- F0 contours were extracted and smoothed using Praat.
- Contours were converted to semitones and their average over time was subtracted to remove global variations (like speaker gender differences).
- All sampled curves were interpolated using a common function basis (B-splines).
- Contours were "registered" in time to a common duration by aligning a specific landmark (the beginning of the final syllable /ri/) to minimize the impact of varying speech rates.
- Functional Principal Component Analysis (FPCA):
- FPCA was applied to the prepared F0 contours. This technique decomposes the variation in the functional data into a set of orthogonal components (Principal Components or PCs).
- Each PC captures a specific mode of variation in the shape of the contours. The first few PCs explain the majority of the variance in the dataset.
- "PC scores" are calculated for each contour, indicating how much that contour deviates from the mean along each PC dimension. These scores are numerical and can be used for further statistical analysis.
Datasets
The study used a dataset of prenuclear pitch accents from a production study on Standard Southern British English (SSBE).
- Source: Recordings from 11 female and 5 male native SSBE speakers reading a narrative text.
- Target: Tokens of the word 'Emory' were extracted.
- Selection Criteria:
- Prenuclear accents (first pitch accent in an intonational phrase, not bearing focus).
- Target word was either the first word or preceded by a two-syllable anacrusis.
- Labeled as rising (L*H) or falling (H*L) according to Gussenhoven's scheme.
- Excluded tokens with initial creak or where modal voicing wasn't maintained for at least 67% of the stressed vowel.
- Final Dataset Size: 126 tokens.
This dataset was not newly created for this paper but was part of a larger ongoing study.
Key Insights and Ideas
- FDA as a Theory-Neutral Tool: FDA provides a way to analyze intonational contours without being overly constrained by pre-existing phonological theories. It allows for the discovery of patterns directly from the data.
- Capturing Global Shape Variation: FPCA effectively captures the primary modes of variation in pitch contours. The first principal component (PC1) explained a significant majority (83.5%) of the variance, indicating it represents the dominant shape characteristic.
- Correlation with Manual Labels: The PC1 scores showed a clear distributional separation between contours manually labeled as "rising" and "falling." This suggests that the primary shape variation captured by PC1 aligns well with established phonological categories, validating the manual labels to some extent.
- Revealing Subtle Phonetic Detail: While PC1 captures the global rise/fall trend, PC2 (and potentially higher PCs) can reveal finer phonetic details. For instance, PC2 might help refine the approximation of contours that don't fit the dominant PC1 pattern, suggesting it captures more localized shape variations.
- Asymmetry in Contour Shape: The analysis highlighted an asymmetry in how pitch contours are realized, beyond just a global rise or fall. The way transitions are made and the shape of pitch movements themselves show variation that might not be apparent in simpler descriptions.
- Potential for Perceptual Studies: The ability of FDA to model and synthesize pitch contours based on PC scores opens avenues for direct perceptual testing of hypothesized phonetic variations.
Key Results
- PC1 Dominance: The first Principal Component (PC1) explained 83.5% of the variance in the F0 contours, representing the main shape variation. PC2 contributed an additional 9%.
- Distributional Separation: PC1 scores for contours labeled as "rising" and "falling" showed distinct distributions, with a clear separation around the zero point.
- Reconstruction Accuracy: Figure 3 demonstrates that considering only PC1 allows for a reasonable approximation of the overall dataset's shape variations, with rising contours generally showing an upward trend and falling contours showing a downward trend after an initial peak.
- Phonetic Detail: The paper suggests that negative PC1 contours tend to exhibit "ampler excursions" than positive ones, and that the way pitch movements are shaped (not just their alignment) differs.
Limitations and Future Work
- PC2 Interpretation: PC2 was noted as being more difficult to interpret and primarily served as a "correction function" to improve the approximation of a small part of the dataset not well accounted for by PC1.
- Meaningfulness of Phonetic Detail: While FDA can identify phonetic variations, the paper acknowledges that it cannot, on its own, determine what is phonologically "meaningful." This requires further investigation, potentially through perceptual studies.
- Perceptual Consequences: The paper suggests that the phonetic details revealed by FDA (like differences in pitch excursion shapes) could have perceptual consequences, citing research on peak vs. plateau contours. However, this is presented as a direction for future work.
- Bridging Theory and Data: The authors propose that FDA can help bridge the gap between phonological theory and phonetic realization, and identify new avenues for research that have been less studied.
- Synthesis and Perceptual Testing: The paper highlights the potential of FDA for synthesizing pitch contours for perceptual testing, which could validate the findings and explore the functional significance of the identified variations.
Summary updated