omzugo/acp-corpus-filtered

ACP Filtered Conversations (via Cortico)

This dataset is a filtered slice of conversation recordings and transcripts from the American Conversation Project (ACP), retrieved via Cortico's ACP integration. "Filtered" means every fragment included here already passed an LLM-based salience pass (the project's internal "wheat vs. chaff" filter) that removed filler, small talk, and interjections — everything kept is a substantive, quote-anchored moment someone actually said.

This is not synthetic data — these are real people's recorded civic conversations. Treat it accordingly (no re-publishing outside this private/gated repo without checking ACP's data-use terms).

What's here

audio/
  conversation-<id>.mp3         Full recording, re-encoded to 64kbps mono
                                 (from the original 192kbps) to keep transfer
                                 size reasonable -- full speech quality, not
                                 the original master. One per conversation
                                 that has audio (205 of 209).
salient_moments_cache/
  conversation-<id>.json         Full filtered result per conversation: every
                                 kept fragment (grouped into "stories"), plus
                                 kept_snippet_indices / cut_snippet_indices
                                 marking exactly what was filtered out.
metadata/
  conversations.jsonl             One row per conversation: Cortico IDs,
                                 title, timing, counts, whether audio exists.
  fragments.jsonl                 One row per KEPT fragment (flattened across
                                 every conversation) -- the easiest starting
                                 point for most uses.

209 conversations total. 205 have audio (4 -- 11127, 11208, 11302, 11313 -- have no recording available on Cortico's side). 11,141 kept fragments across the corpus.

How the audio relates to the transcript

Audio is not clipped to individual fragments -- each audio/*.mp3 is the complete recording (re-encoded to 64kbps mono for size, not otherwise cut or altered). Every fragment instead carries its own audio_start_offset / audio_end_offset (seconds into that same file).

This mirrors how the source app itself plays these back: it plays the full track continuously and uses these same offsets to highlight/seek through only the kept spans, silently skipping everything that was filtered out, rather than pre-cutting the audio into separate clips. Use fragments.jsonl's audio_file + audio_start_offset/audio_end_offset to do the same -- whether that means seeking during playback, or trimming clips yourself if your use case wants that instead.

metadata/fragments.jsonl schema

field	meaning
fragment_id	<conversation_id>:<story_id>:<index> -- stable, unique per fragment
conversation_id	This repo's filename stem, e.g. conversation-10900
cortico_conversation_id	Cortico's own numeric conversation ID (int)
conversation_title	Cortico conversation title
story_id	Groups fragments that are part of the same narrative thread
speaker_id / speaker_name / speaker_real_name	Speaker identity (diarized; real name where the speaker stated it in-conversation)
bio_context	Short bio blurb inferred for this speaker/story
phase	Rough position in the conversation (e.g. intro)
text	The fragment's verbatim text
char_start / char_end	Character offsets into the conversation's cleaned transcript
audio_start_offset / audio_end_offset	Seconds into audio_file
reasoning	The model's own reasoning for why this fragment was kept
has_audio / audio_file	Whether/where the matching audio lives

metadata/conversations.jsonl carries the conversation-level Cortico provenance (cortico_conversation_id, cortico_forum_id, cortico_host_id, title, start_time, num_participants) for every conversation, whether or not it has audio.

Provenance

Source: American Conversation Project (ACP), via Cortico's API (org 353). Transcribed by Cortico (machine transcription); salience filtering and narrative/story grouping performed by an LLM pipeline (backend/salient_extractor.py in the source project), with every kept fragment grounded back to a verbatim quote in the original transcript.