PDF-to-Audio Accessibility BenchmarkA reproducible benchmark for the extraction stage behind VoiceBrief narration. The July 30, 2026 strict synthetic rerun passed 7 of 7 cases, including scanned-page OCR and an interleaved two-column content stream.Extraction source revision c343c1cbfdb5. Download the raw JSON results released under CC BY 4.0. The earlier July 28 result remains available as historical evidence.Strict resultAll seven strict synthetic cases pass: clean text, logical and interleaved columns, scanned-page OCR, repeated page furniture removal, table spacing, and line-break hyphenation.Method

The runner generates seven VoiceBrief-owned synthetic PDFs, including a scanned image-only page and multiple two-column layouts.

Each file goes through the same standard extraction function used for PDF uploads. OCR runs only when standard extraction returns no usable text.

A case passes only when every expected anchor is recovered, adjacent anchor pairs appear in the intended order, and repeated-furniture limits are respected.