Hypline¶
Hypline is a command-line toolbox for analyzing fMRI hyperscanning data collected during dyadic conversation. These interactions are complex and dynamic, and analyzing them often requires coordinating multiple steps across behavioral and neuroimaging data. Hypline provides a standardized workflow for structuring hyperscanning datasets, preparing and denoising fMRI data, and running analyses.
Hypline currently supports encoding analyses. Over time, we aim to expand the toolbox, with development from our team and yours, to support additional approaches including hyperalignment and shared response modeling, intersubject correlation, mental state decoding, hyper-hidden Markov modeling, natural language processing, and more. The pipeline provides defaults developed through extensive testing while allowing researchers to customize individual analysis choices.
What is an encoding model? An encoding model learns a mapping from features of a stimulus or behavior to measured brain activity, then tests how well those features predict responses in held-out fMRI data. Encoding models are widely used in naturalistic fMRI to study what information is represented across the brain during continuous experiences such as listening to speech, watching movies, or engaging in conversation. In conversational hyperscanning, encoding models allow researchers to relate features of an ongoing conversation — such as its phonemic, semantic, syntactic, or acoustic content — to the BOLD responses of the people producing and comprehending that speech. For example, Zada et al. (2026)1 used encoding models to examine shared neural systems involved in speech production and comprehension during real-time dyadic conversation. To learn more about how encoding models work, see How the encoding model works.
Hypline supports the main steps needed to run an encoding analysis on dyadic conversation data: transcribing recorded speech, generating stimulus features and their confounds, denoising fMRIPrep BOLD data, and fitting and evaluating encoding models. Its commands are modular: each performs one step and can be run independently, while all commands operate within the same BIDS-style dataset.
Where to go next¶
- Need to install? See Installation.
- Want to try it now? Follow the Tutorial for a complete run on a downloadable example dataset.
- New to hypline? Start with Hypline dataset layout to learn how a dataset is organized.
- Want command details? See the User guide for each command's arguments and options.
- Want to understand the analysis better? How the encoding model works and Feature families cover what you can predict and how the analysis is set up.
Citing Hypline¶
If you use Hypline in your research, please cite it:
- Open Hypline's Zenodo record.
- Pick the version you used under Versions (the latest is shown by default).
- Copy the citation from Citation, in the format you need (APA, BibTeX, and others).
If you use the encoding model, please also cite Zada et al. (2026)1.
-
Zada, Z., Nastase, S. A., Speer, S., Mwilambwe-Tshilobo, L., Tsoi, L., Burns, S. M., Falk, E., Hasson, U., & Tamir, D. I. (2026). Linguistic coupling between neural systems for speech production and comprehension during real-time dyadic conversations. Neuron, 114, 1–14. https://doi.org/10.1016/j.neuron.2025.11.004 ↩↩