hbmartin/podcast-transcript-archive

Podcast Transcripts

Downloaded podcast transcripts, with original files and extracted text. Generated with overcast-to-sqlite on 2026-10-13T21:13:00.492425+00:00. This snapshot contains 11941 archived files, of which 11929 have extracted text and 12 have extraction errors. 11733 rows have original source URLs.

Contents and schema

data/train.parquet contains one row per archived file. raw-transcripts.tar.gz preserves all original files under raw/, including their directory structure. Multiple files can describe the same episode, and identical content can occur in different rows. train is a single collection, not a curated training/evaluation split.

Column	Meaning
id	SHA-256 of the relative archive path; stable for a given path
podcast / episode	SQLite titles when matched, otherwise directory/filename
text	Extracted UTF-8 text; null when extraction fails
source_url	Original transcript URL when available
feed_url / episode_url	RSS feed and episode page URLs when available
published_at	Publication date as recorded in the feed, without date normalization
format	Detected content format; extension fallback for extraction failures
raw_path	Original file's path inside raw-transcripts.tar.gz
sha256	SHA-256 of the original file bytes; usable for deduplication
extraction_status	ok or error
extraction_method	Text extraction strategy, including HTML selection strategy
extraction_error	Error description when extraction fails

Processing and limitations

Formats are detected from content and filename hints, including subtitle content stored as .bin and PDFs stored as .octet-stream. Subtitle indices and timing lines are removed; WebVTT voice labels are retained when present. JSON full text or segments and Podlove XML items are read. PDFs use embedded text extraction without OCR. HTML scripts and common navigation elements are removed; transcript containers are preferred, then article/main content, then the remaining body. html_article and html_body rows can include episode notes or other page text. Plain text preserves speaker labels and any timing markers already in its text. Whitespace is normalized. No translation, content deduplication, or transcript accuracy correction is performed. Failed extractions remain as rows with null text, and their original files remain available. See extraction-report.json.

Metadata uses an explicit selection of feed/episode fields from a read-only SQLite connection. Source URLs and dates are null when no matching metadata is available.

Loading

from datasets import load_dataset

dataset = load_dataset("hbmartin/podcast-transcript-archive", split="train")
readable = dataset.filter(lambda row: row["extraction_status"] == "ok")

To download and unpack all original files:

uvx --from huggingface_hub hf download hbmartin/podcast-transcript-archive raw-transcripts.tar.gz \
  --repo-type dataset --local-dir .
tar -xzf raw-transcripts.tar.gz

Each row's raw_path points to its file in the extracted raw/ directory. The bundle avoids thousands of individual file commits when publishing updates.

Regeneration

The standalone generator and its dependency declarations are included in scripts/generate_transcript_dataset.py. With the original archive and optional Overcast SQLite database available, run:

uv run scripts/generate_transcript_dataset.py archive/transcripts \
  --database overcast.db --output archive/huggingface/podcast-transcripts \
  --repo-id hbmartin/podcast-transcript-archive

Generation is local. Upload the generated output directory to this dataset repo using hf upload --repo-type dataset.

Licensing and provenance

Transcripts originate from the podcasts identified in each row. Copyright and any applicable licenses remain with their respective owners; this collection does not grant a blanket license to the transcript content. The generator code is covered by the project's Apache-2.0 license, included as scripts/LICENSE.