Hypline

Hypline is a command-line toolbox for cleaning and analyzing data from hyperscanning studies involving dyadic conversations. Its commands are modular: each does one job — transcribe audio, generate features, denoise fMRIPrep BOLD, fit an encoding model — and runs on its own, all inside one BIDS-style dataset.

An encoding model predicts the brain's BOLD response from features of the speech a participant heard. Hypline prepares both sides of that fit — the stimulus features and the denoised BOLD — then fits and scores the model with encoding. You bring two inputs: the stimulus audio and the fMRIPrep-preprocessed BOLD. Hypline starts from fMRIPrep's outputs; it does not preprocess the BOLD itself.

Hypline implements the encoding-model approach of Zada et al. (2026), Neuron (10.1016/j.neuron.2025.11.004), which used fMRI hyperscanning and language-model features to study the shared neural systems for speech production and comprehension in real-time dyadic conversations.

Installation

pip install hypline

Also installable with uv (uv add hypline) or poetry (poetry add hypline). This installs the hypline command:

hypline --help

[!NOTE] hypline transcribe decodes audio through FFmpeg, which must be installed separately and on your PATH. Other commands do not need it.

The pipeline

Hypline's commands compose into a pipeline. Every command reads from a shared dataset root and writes back into the same tree. Most fall into two independent branches — a stimulus branch and an fMRIPrep branch — that prepare the two sides the encoding branch then joins:

featuregen phonemic also generates the matching phonemic confounds by default, so you rarely call confoundgen phonemic directly. And you needn't run every step: any command works on its own once its inputs exist — run transcribe alone for transcripts, or denoise alone to clean fMRIPrep BOLD.

Quick start

Once your files sit where hypline expects (see the dataset layout), you only ever point a command at the dataset root — it finds its own inputs from there, so you never pass file paths. End to end, the whole pipeline is four commands:

# stimulus branch: audio → transcripts → features (+ phonemic confounds, auto)
hypline transcribe data/ --audio-ext .wav
hypline featuregen phonemic data/

# fMRIPrep branch: clean the BOLD with a motion + drift model, read straight
# from fMRIPrep's confounds table
hypline denoise data/ \
  --columns trans_x,trans_y,trans_z,rot_x,rot_y,rot_z,cosine

# encoding branch: fit the model that maps features onto the denoised BOLD
hypline encoding train data/ \
  --tasks conv \
  --features phonemic \
  --desc v1 \
  --fold-by none

After this, data/ holds phonemic features plus desc-denoised BOLD — the two sides the encoding model needs — and a fitted model under results/.

Documentation

Full guides and per-command reference live at the project documentation. New to hypline? Walk through a full run on the example dataset, or read The hypline dataset layout — every command depends on it. To load a fitted model or eval back into Python, see the encoding results API.

License

Released under the MIT License.