qb-audio — read-aloud audio for quizbowl questions

TTS read-aloud recordings of quizbowl questions from qbreader, one Opus file per tossup / bonus. Built so any quizbowl app can add listening practice over plain HTTP — no install, no API key. Generation is in progress (July 2026); the dataset grows as the run proceeds.

Layout

• tossups/{qid[-2:]}/{qid}.opus, bonuses/{qid[-2:]}/{qid}.opus — audio, Opus 24 kHz mono. qid is the question's qbreader _id; files shard by its last two hex characters.
• tossups/{qid[-2:]}/{qid}.json, bonuses/... — chunk-offset sidecar: {"v": 1, "chunks": [[start_s, end_s], ...], "texts": ["chunk text", ...]}. Synthesis is per-chunk, so cumulative chunk durations map audio time to text position (for buzz points, reveal pacing).
• audio_index.json — manifest: {"tossups": [qid, ...], "bonuses": [qid, ...]}, the questions that have audio. Fetch this first and restrict your queue to it.

The spoken text is a cleaned form of the question text (pronunciation guides applied, markup and moderator directives stripped) — the sidecar texts are exactly what the audio says.

How it was made

Synthesized with Chatterbox TTS (MIT, Resemble AI) on a single RTX 4090, with an ASR verification gate (OpenAI Whisper, MIT) that re-rolls clipped or runaway generations. Pipeline code is open source under the qbsuite org (MIT).

Credits and licensing

• Question text was written by the packet authors and editors of the respective tournaments and reaches this dataset via qbreader, which distributes the community question archive. The audio is a synthesized reading of that text; rights in the underlying questions remain with their authors.
• The dataset is shared for study and practice in line with quizbowl community norms, the same basis on which the text archive circulates. It is not licensed for uses beyond that (the other license tag above reflects this — there is no blanket open-content license to grant).
• Audio synthesis: Chatterbox TTS (MIT © Resemble AI). QA: Whisper (MIT © OpenAI).

If you are an author or editor of included questions and want your material removed, open an issue on any qbsuite repo or use the dataset's community tab and it will be taken down.