scotus

Larry SCOTUS Oral Argument Aligned Segments

Built from thelarryproject/scotus-unaligned.

U.S. Supreme Court oral arguments, October Terms 2010–2025, cut into short single-speaker clips with matching transcript text. Each argument's transcript was aligned to its audio, then split at speaker changes, long pauses, and sentence ends into clips of at most 25 seconds. Contains 395,983 clips (~1,191 hours) from all 1,050 argument sessions, un-pruned.

Fields

• docket, term, case_name, date_argued, pdf_url, page_url: copied from the argument the clip came from
• audio: Arrow audio typed column of (bytes, path); 16 kHz mono FLAC, path is empty
• text_original: the transcript text for the clip, as written by the court reporter
• text_normalized: the same text spelled out as spoken, in uppercase letters, apostrophes, and spaces (e.g. "Mr." -> MISTER, "09-400" -> NINE FOUR HUNDRED)
• speaker_id: a consistent speaker id across arguments, e.g. scotus:JUSTICE KAGAN, scotus:CHIEF JUSTICE ROBERTS, scotus:adv:PAUL CLEMENT
• start_s, end_s: where the clip sits in the original argument audio, in seconds
• duration_s: clip length in seconds
• avg_word_score: how confidently the clip's words were aligned (closer to 0 is better)
• ctc_mismatch: how poorly the clip's audio matches its text when scored on its own (lower is better); empty when the clip is too short to score

Construction

Aligned with facebook/wav2vec2-large-960h-lv60-self. See provenance.json for the producing commit, timestamp, and ancestry.

Known corrections

From the manifest

• 13-7120rearg: the court's page links the original argument's transcript and date; this has been manually corrected to the reargument transcript and 2015-04-20

From the unaligned dataset

• Minor misplacements of speaker labels in the original transcripts

From this dataset

• Speaker labels are mapped to one id per person: justices are matched against the bench sitting on the argument date, and advocates against each transcript's list of appearances
• Reporter typos in names are corrected, e.g. "ERICJ FEIGIN" -> ERIC FEIGIN, "JEFFERY FISHER" -> JEFFREY FISHER, "IRV GORNSTEIN" -> IRVING GORNSTEIN, "STUART DUNCAN" -> KYLE DUNCAN; a justice mislabeled as an advocate (e.g. "MR. KAGAN") is attributed to the justice
• Stage directions such as "(Laughter.)" and "(Pause.)" are removed from the text
• Where text can be spoken more than one way (numbers, years, "v." as "V" or "VERSUS"), the reading that best matches the audio is used

Known gaps

• Clips under ~2 seconds are mostly quick interruptions where speakers overlap; many have high ctc_mismatch and may contain parts of the other speaker
• Neighboring clips can overlap by up to 0.2 seconds, so a clip's end can include the first sound of the next one
• 9 clips have speaker_id scotus:unknown, mostly the Chief Justice's opening and closing lines in a few 2013–2014 arguments
• No quality, duration, or ctc_mismatch filtering is applied here; that is left to downstream consumers

The audio and transcripts are Public Domain, but this redistribution is CC-BY-4.0 under the larry project.

Preprocessing pipeline code is at https://github.com/Mekadrom/larry, and this dataset is fully reproducible with a single command in that repo:

uv run preprocessing --config_file configs/preprocessing/voice/aligned/thelarryproject_scotus-aligned.yaml --output_dir ~/.cache/larry/datasets/voice/aligned/thelarryproject_scotus-aligned --clean