Bring different media into one processing flow.

Build routing, outputs, and failures into the workflow.

Choose the path from the source

Prefer usable PDF text layers and embedded video subtitles; fall back to OCR or speech recognition when needed.

Keep one output contract

Each material produces content.md, meta.json, and raw/. Batch results appear in index.md for both programmatic and human use.

Make exceptions explicit states

Distinguish ok, partial, no_speech, and failed. Low-confidence OCR retains a review list; images with little text are marked for visual review.

Two test materials, two processing paths.

These excerpts come from test fixtures processed on October 26, 2026, not client materials. The Chinese subtitle is translated below for readability; linked outputs retain the actual original text.

Video with subtitles

5-second test video → embedded subtitle extraction → 38 characters in two timestamped segments. Source: subtitle; no speech recognition used for this run.

[0:00] This is the first embedded-subtitle test sentence. [Translation]
[0:02] This is subtitle track line two

View the actual subtitle output (original text) ↗

PDF with a text layer

1-page PDF → direct text-layer extraction → 174 characters. Text is organized by page and labeled text-layer.

media-extract text layer test document

View the actual PDF output ↗

The downloadable copies only remove the machine's absolute source path; text and processing provenance remain unchanged.

Test each branch; do not equate test counts with accuracy.

Verified paths

Checked embedded video subtitles, PDF text extraction, and output source/status fields. Scanned-PDF OCR passed with normal local-machine permissions.

Test coverage

Full suite rerun on October 26, 2026: 47 passed, zero failures or skips, 15.321 seconds. Covers text processing, naming, PDF, subtitles, speech recognition, and resumable batches. These are fixture-based workflow checks, not general recognition accuracy. View verification notes (Chinese) ↗

Limitations

OCR and speech recognition can misread names and require source review. No production usage, accuracy, or time-saving claims are made. No public upload service is provided.