SpeechGraph

SpeechGraph is a computational pipeline for analyzing the structure of transcribed speech through directed lexical graphs. It extracts graph-theoretic and recurrence-based features from discourse and evaluates their association with impulsivity dimensions measured by the Barratt Impulsiveness Scale (BIS-11).

Pipeline

1. Parsing of orthographic transcripts into token sequences.
2. Graph construction: directed lexical transition graphs where each node is a unique token and each directed edge represents a transition between consecutive tokens within the same discourse segment.
3. Sliding-window metrics quantifying topology, connectivity, recurrence, and path structure of the graph.
4. Z-scores via permutation testing, comparing observed metrics against a null distribution generated by shuffling tokens within discourse segments.
5. Spearman correlations (simple and partial) between graph metrics and impulsivity dimensions.
6. Predictive regression using Monte Carlo cross-validation to assess how well graph metrics predict impulsivity scores.
7. Hyperparameter optimization with Optuna, exploring multiple regression algorithms and RFE-based feature selection.

Graph representation

$$ G = (V, E) $$

where each vertex $v \in V$ is a unique lexical token and each directed edge $(v_i, v_{i+1}) \in E$ represents a transition between consecutive tokens within the same segment. Edges are weighted by their occurrence frequency:

$$ w(v_i, v_j) = \text{number of times the transition } v_i \rightarrow v_j \text{ occurs} $$

Z-score via permutation

For an observed metric $x_{\text{obs}}$ and its null distribution ${x^{(r)}}_{r=1}^{R}$ generated by within-segment permutation:

$$ z_x = \frac{x_{\text{obs}} - \mu_{\text{rand}}}{\sigma_{\text{rand}}}, \qquad \mu_{\text{rand}} = \frac{1}{R}\sum_{r=1}^{R} x^{(r)}, \qquad \sigma_{\text{rand}}^2 = \frac{1}{R}\sum_{r=1}^{R} \left(x^{(r)} - \mu_{\text{rand}}\right)^2 $$

Regression

For each target $y$ and design matrix $\mathbf{X}$ (selected features):

$$ \hat{y} = \beta_0 + \mathbf{X}\boldsymbol{\beta} $$

Performance is evaluated via:

$$ \text{MAE} = \frac{1}{n}\sum_{i=1}^{n} |y_i - \hat{y}_i|, \qquad R^2 = 1 - \frac{\sum (y_i - \hat{y}_i)^2}{\sum (y_i - \bar{y})^2} $$

Interactive dashboard

All results —cross-experiment comparisons, distributions, feature analysis, optimization, and SHAP analysis— are available on the live dashboard:

➡️ SpeechGraph Dashboard