To turn a PDF article into useful audio, first prepare a reading script that preserves the article's order and argument. Then use a program with a documented audio export, such as Balabolka on Windows or the ElevenLabs Audiobooks web workspace. Two-column layouts, references and tables need editorial decisions before they become speech.
This guide uses provider documentation reviewed on 6 November 2026. It does not report a conversion test of an academic paper, an OCR accuracy score or a comparison of voice quality.
Check the reading order of the article
Start with one difficult page rather than importing the entire paper and trusting the extracted text. Copy a paragraph into a text editor and compare its order with the visible PDF. Then check a paragraph that crosses a page boundary.
Look for these specific failures:
- Lines from the left and right columns alternating inside one sentence.
- A figure caption inserted halfway through a body paragraph.
- The last line of a page separated from its continuation by a header or page number.
- Superscript citations, minus signs, units or symbols missing from the copied text.
- Words split by layout hyphenation that have not rejoined correctly.
PDF text being selectable does not prove that its reading order is correct. Adobe's accessibility guidance explains the role of searchable text and document structure tags. A scan with no useful text needs OCR first; a PDF with text in the wrong order needs a different extraction or editing approach.
If Balabolka's extraction is poor, its official FAQ documents using Xpdf's pdftotext through Custom Text Import. That is an alternative extractor, not a guarantee that every two-column paper will be correct. For a short article, manually assembling and checking the body text can be easier to audit than repeatedly changing extraction tools.
Decide how to handle citations and tables
Keep the original article alongside a separate narration script. Make omissions explicit in your working notes. Do not let an application's cleanup setting decide whether a footnote is essential to the argument.
| Article element | A workable treatment | What to preserve |
|---|---|---|
| Inline citations | Keep them for close study, or move them to a separate references segment for uninterrupted listening. | The link between a claim and its source. Do not silently make quoted work sound like the author's own result. |
| Footnotes | Read essential qualifications after the sentence; move optional detail to a clearly named note section. | Exceptions and definitions that change the meaning. |
| Tables and figures | Retain the caption and a reference to the original page. If you write a spoken explanation, check it against the values and label it as your explanation. | Units, axes, row/column relationships and the original evidence. |
| Equations | Keep an accurate verbal reading if you can verify it; otherwise mark the place to consult the PDF. | The equation number and page, instead of a guessed interpretation. |
| Headers and bibliography | Remove repeated page furniture; keep the bibliography separately if you skip it in the main narration. | Article identity and a way to recover its references. |
These are editing choices, not claims that a speech engine automatically understands figures or research. The header and footnote cleanup guide gives the separate removal workflow. If you need a faithful full reading, avoid asking a summarizer to rewrite the article. A generated discussion can be useful for another purpose, but it is a different output from an audiobook.
Create and review a short sample
Choose the program after the text is ready:
For Windows and an installed voice: Balabolka documents PDF text extraction and saving speech to an audio file. Open the checked text, choose a voice and rate, and save a short section. Its FAQ covers named bookmarks and chapter output if you want the article's sections separated. See the Windows workflow for that route.
For a web editor with a single narrator: ElevenLabs Audiobooks documents manuscript upload or paste, a Single cast option, selectable sections and MP3/WAV export. Import the article or paste the edited script, inspect the parsed sections, choose one narrator, and generate a representative sample. Export only the intended scope. A single file or chapter ZIP is available in the documented workflow, and ungenerated sections can consume credits during export.
For a scan in NaturalReader Personal: the input documentation describes OCR ranges of up to 50 pages at a time. That is separate from the audio export limit: MP3 generation requires a subscription and an eligible voice, with up to 20 document pages per conversion. Free accounts and Free Voices do not provide that MP3 route. Verify your account entitlement before processing the article.
Use a sample containing a column transition, a citation and a number with a unit. Follow it against the source sentence by sentence. A natural-sounding voice can still read an extraction error confidently. Stop and repair reordered paragraphs, changed quantities or missing qualifications before generating the rest. The preview guide gives a reusable listening checklist.
Keep a route back to the original PDF
Save the source PDF, corrected script and audio in the same project folder. Use the article title and version in the filenames so that an updated manuscript does not become confused with an earlier recording.
Keep a small manual index. For example:
01-abstract.mp3 PDF page 1, Abstract
02-methods.mp3 PDF pages 2–4, Methods
03-results.mp3 PDF pages 4–6, Results; consult Figure 2 on page 5
references.txt Original reference list, retained for lookup
This is an illustrative naming scheme, not an output generated by any named product. Record the page numbering convention when printed page numbers differ from the PDF viewer's count. If you omit a table or move a note, record that decision at the corresponding section.
Finally, test seeking and resuming in the player you will use, then check the last sentence against the article. Keep chapter-to-page notes even if the player remembers your position: a playback timestamp alone is a poor citation.
The PDF workspace can extract and edit text for review. It does not interpret research figures or provide synchronized word-by-word PDF playback. A clear script and source index remain useful whichever export tool you choose.