Smoothing Spline ANOVA for Speech Contour Analysis

24 concepts22 papers1 proposal

When we speak, our vocal tracts are in constant, fluid motion. If you capture an ultrasound image of a tongue during speech, you do not see a single static point; you see a complex, continuous curve shifting over time. Similarly, the acoustic output of speech—such as the pitch ($F_0$) or formant frequencies—unfolds as a continuous trajectory rather than a single note.

Traditionally, phonetic analysis has dealt with this complexity by reducing curves to single numbers. Researchers might measure the maximum height of the tongue dorsum, or extract the pitch at exactly the midpoint of a vowel. While simple, this reductionist approach discards crucial spatial and temporal dynamics.

To capture the true nature of speech, linguists and speech scientists increasingly rely on Smoothing Spline Analysis of Variance (SSANOVA). This statistical framework allows researchers to compare entire curves—whether they represent physical shapes in the mouth or acoustic patterns over time—without losing localized details.


The Intuition: Moving Beyond Point Measurements

Imagine trying to determine if two speakers have different strategies for pronouncing a vowel. If you only measure the tongue at its highest point, you might conclude their pronunciations are identical. However, one speaker might be advancing their tongue root, while the other is raising the blade of their tongue.

SSANOVA solves this by treating the entire contour as the unit of analysis. The method works in two main steps:

  1. Spline Fitting: It takes noisy, discrete data points (such as coordinates traced from an ultrasound frame or pitch values sampled every few milliseconds) and fits a smooth mathematical curve called a smoothing spline. This curve balances accuracy to the data points with a penalty for excessive wiggliness.
  2. Analysis of Variance: Just as a traditional ANOVA tests whether the means of different groups are statistically distinct, SSANOVA decomposes the variance of the functional curves into main effects and interactions based on experimental conditions.

The real power of SSANOVA lies in its visual and mathematical output: confidence intervals that span the entire curve. When comparing two conditions, the model generates a representative curve for each, flanked by a confidence band. Wherever these bands do not overlap, the difference between the conditions is statistically significant. Instead of asking if two speech sounds are different, researchers can pinpoint exactly where or when they differ.


Mapping the Geography of the Tongue

The most prominent application of SSANOVA is in articulatory phonetics, particularly for analyzing data from ultrasound tongue imaging (UTI) and real-time magnetic resonance imaging (rt-MRI).

In ultrasound studies, the edge of the tongue is tracked frame-by-frame, yielding a sequence of mid-sagittal coordinates. Lilienthal, in a study on Scottish English, demonstrated how SSANOVA could map the precise regions where the tongue shape shifts before an /r/ sound, revealing subtle articulatory adjustments that were not always visible in the acoustic data.

Because tongue shapes vary wildly between individuals due to unique anatomy, SSANOVA is highly effective at isolating systematic linguistic differences from anatomical noise. As Gili Fivela et al. noted, the method quantifies variations by accounting for the overall form, rotation, and translation of the tongue contours. This allows for robust comparisons across different prosodic boundaries or speaker groups. For instance, Alowonou et al. used SSANOVA to construct representative curves that clearly distinguished the higher and more forward tongue configurations of male Ewe speakers compared to female speakers.

Traditional Analysis:  [Point A] vs [Point B]  --> Loses shape context
SSANOVA Analysis:      [=== Curve A ===]       --> Compares entire shape;
                       [=== Curve B ===]           identifies regions of non-overlap

This regional precision helps resolve long-standing debates in phonological theory, such as whether certain sound changes represent a complete merger of categories or if speakers maintain a "covert distinction." Archangeli et al. applied SSANOVA to Scottish Gaelic lenition, finding that speakers maintained subtle, statistically significant distinctions in specific regions of the tongue contour, challenging the idea of complete neutralization. Similarly, Sung leveraged the technique to track individual variations in Korean palatalization gestures.


Tracking the Dynamics of Sound

The utility of SSANOVA extends seamlessly from spatial curves (like the tongue) to temporal curves (like acoustics). Speech sounds are inherently dynamic; a vowel is not a single static frequency, but a trajectory that transitions from the preceding consonant and moves toward the next.

McAuliffe et al. utilized SSANOVA to analyze the entire acoustic trajectory of vowel formants in spontaneous speech. By modeling formants as continuous curves over their entire duration, the authors showed how word predictability alters the dynamics of vowel production, a finding that would be obscured by single midpoint measurements.

The framework is equally powerful for prosodic analysis, where researchers study intonation and tone through fundamental frequency ($F_0$) contours:

  • Sentence Intonation: Hsu et al. used SSANOVA to visualize how Taiwan Mandarin speakers manipulate pitch contours over time to distinguish between questions and statements when using ambiguous "wh-words."
  • Tone Interactions: Liu et al. applied the method to Taiwanese Southern Min to track how neutral tones dynamically shift their pitch targets based on the preceding consonant and vowel context.
  • Clinical Linguistics: Liu et al. mapped $F_0$ contours in Mandarin speakers with Parkinson's disease, using the non-overlapping confidence intervals of SSANOVA to illustrate exactly where patients fail to modulate their pitch compared to healthy controls.

Why the Idea Works

SSANOVA succeeds because it respects the continuous nature of speech data while maintaining statistical rigor. If a researcher were to perform a t-test at 100 individual points along a tongue curve or a pitch track, they would run into the "multiple comparisons problem," drastically increasing the likelihood of false positives.

By treating the curve as a single function and using smoothing splines, SSANOVA inherently handles the dependency between adjacent points. It provides a mathematically sound way to answer a fundamental question in linguistics: Where exactly does a sound change make a difference?

Key Foundational Work

The articulatory and acoustic impact of scottish English /r/ on the preceding vowel-onset (2009) — contributed 2 concept elements to this direction.
An analysis of vocal tract shaping in English sibilant fricatives using real-time magnetic resonance imaging (2008) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_smoothing_spline_anova_8d754fea,
  title = {smoothing spline ANOVA},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/smoothing-spline-anova/rdr_8d754fea920946789b4b76825f082b8a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_smoothing_spline_anova_8d754fea_v1,
  title = {smoothing spline ANOVA},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/smoothing-spline-anova/rdr_8d754fea920946789b4b76825f082b8a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake