This is the HuggingFace organization for podscripter, a Dockerized local-first transcription tool built on OpenAI Whisper, pyannote.audio speaker diarization, and sentence-transformers punctuation restoration. Primary language focus: English, Spanish, French.
This org doesn't publish models — Whisper and pyannote live in their own upstream orgs. What lives here is the supporting data that the podscripter project owns and republishes under permissive licenses, primarily for testing and reproducibility.
podscripter-project/test-fixtures
— small, curated EN/ES/FR audio clips (CC-BY 4.0) used by podscripter's Tier 1 regression
tests. Audio is sourced from permissively licensed public corpora (LibriSpeech, FLEURS,
Common Voice, VoxPopuli, AMI, MLS) and trimmed/concatenated to exercise specific
pipeline code paths (single-speaker ASR, multi-speaker diarization, chunked-mode
transcription). Each clip ships with verbatim transcripts, speaker turns, source
attribution, and per-fixture WER/DER thresholds.Everything published here is permissively licensed (CC-BY 4.0 or CC0 1.0). Aggregate
licenses match the most restrictive component — typically CC-BY 4.0, which requires
attribution and indication of changes when redistributed. Per-source attribution lives in
each artifact's dataset card and (for the test-fixtures) in
tests/fixtures/audio/LICENSES.md
in the podscripter repo.
NC/ND-licensed sources are deliberately excluded so artifacts here can be freely redistributed.
Issues, fixture proposals, and bug-reproduction clips all go through the
podscripter GitHub repo. The
contribution workflow for new audio fixtures
covers trimming, licensing requirements, the .expected.json schema, and bumping
HF_REVISION so the dataset and tests stay in lockstep.