Glossary

Hypline gives a handful of ordinary words a specific meaning. This page defines them in one place; the rest of the docs link here rather than re-explaining. Terms are grouped by where you meet them.

Dataset structure

Dataset root : The single directory every command takes as its argument. Hypline finds all inputs and writes all outputs by following a fixed layout beneath it, so you never pass file paths. See The hypline dataset layout.

Dyad : The pair of subjects who held one conversation while both were scanned. Hypline is a hyperscanning pipeline, so the conversation belongs to the dyad, not to either partner.

Sub-keyed / dyad-keyed : Which identity a file leads with. A sub-keyed file is derived from one brain (raw BOLD, denoised BOLD, a fitted model); a dyad-keyed file is derived from the shared conversation (audio, transcripts, features, confounds). The two are bridged through participants.tsv. See Subject vs. dyad.

Segment : A named time window within a run (a trial, block, or condition), declared in the run's events.tsv. Segments are what let you generate per-trial features and filter down to specific conditions. See Segments and metadata.

desc variant : A named alternative derivation of the same source, tagged with --desc and kept in its own subdirectory so variants sit side by side. Two phonemic-confound flavors (phonemic-onset, phonemic-rate) are desc variants of one kind.

Features

Feature : A stimulus-derived predictor for the encoding model — a per-word or per-TR vector describing the speech a participant heard. The four families are phonemic, semantic, spectral, and syntactic; see Feature families.

Confound : A stimulus-derived nuisance regressor (speech onset, speech rate) partialled out inside the encoding fit, so a feature cannot claim signal a confound explains. Distinct from denoising nuisance, which is removed from the BOLD beforehand.

Downsample : Binning a per-word feature onto the BOLD TR grid, since features are timed to words but BOLD is sampled per TR. Controlled by --downsample (mean or sum).

The encoding model

Encoding model : A model that predicts a voxel's BOLD signal from a weighted sum of speech features. Hypline fits one per subject with encoding train.

Band : One part of a banded-ridge model with its own regularization strength. Each feature is a band, all confounds share one band, and a reserved task band absorbs the production-versus-comprehension signal offset. Also the axis an eval scores on. See How the encoding model works.

Delays : The set of time lags (in TRs) at which each feature enters the model, since a voxel's response to a word is spread over the following seconds. Set with --delays.

Split : Fitting separate weights for when the subject is producing speech and when they are comprehending, by duplicating each feature into a speaking and a listening copy within its band. On by default; --no-split fits one shared set.

Analysis

Source / model / target : The three independent subject roles in encoding analyze. Source drives the prediction's inputs (whose speech), model supplies the weights (whose trained model), target is the brain being predicted (whose BOLD). See Choosing source and model.

Within-brain analysis : Scoring a subject's own model, driven by their own speech, against their own brain (source: self, model: self). The baseline case.

Cross-brain analysis : Predicting one partner's brain using the shared conversation and the other partner's model — the analysis hypline is built for. It comes in a partner-driven form (source: partner, model: partner) and a self-driven form (source: self, model: partner).

Pseudo-dyad : A baseline where a subject's own speech and brain are scored with a model trained on someone they never conversed with. Because source and target are still the same subject, it runs without warning; the mismatch is in the model.

Scramble control : A control where source and target belong to different dyads, so one conversation's speech is paired against another's brain, aligned only by matching run rather than a shared conversation. Mechanically valid but not a fit; analyze warns when it detects this.

Role (prod / comp / both) : The turn subset a score covers, from the target's own turns: prod (target speaking), comp (target listening), both (either). An eval reports every band's score for each role. See Reading an encoding result.

Eval : The output of analyze — a netCDF file of per-voxel scores indexed by fold, band, role, and voxel, loaded back with load_eval. See Reading an encoding result.

Split score : The value stored in an eval: one band's own share of the joint model's correlation, from himalaya. A decomposition, so the bands sum to the joint score — but a single band's share is not itself a Pearson correlation and can be negative.