mdonigian/scotus-vocal-data

Mimir: Supreme Court voice and vote data

Mimir studies whether measurements of justices' speech during oral argument help predict their votes. Start with the evaluated model and its model card:

• Mimir vote predictor: the selected model, portable inference, final evaluation, coverage and limitations.
• Complete supporting model data: fitting data, separate final inputs and targets, candidate models, predictions, acoustic measurements, new word and diarization outputs, source inventories and validation receipts.
• Consolidated source and acoustic release: original Oyez audio and annotations, historical acoustic products and reusable processing outputs. Start at current/recordings.parquet for the historical acoustic catalog and consult its dataset card for schemas and missing measurements.

The model and its supporting data are published on normal main branches. The source archive uses the normal versioned branch release-2026-09.1. Published model metadata and release manifests pin immutable commits and content hashes for reproducible downloads. Identical source audio is referenced rather than duplicated in each model bundle.

Original sources are preserved. Unavailable measurements and fallback predictions are explicit. The model card explains what the final comparison establishes, including the distinction between vocal delivery and speaking amount. Prediction before Oyez transcripts arrive has not been evaluated by this model.

Earlier pitch and richer-feature benchmarks remain available as research history. Use the current model card to interpret the later-term evaluation.

Code, methods and research evidence are maintained in the project repository. Credentials and deleted human reviews are excluded from these releases.