Table of Contents
On every ebook forum, one class of complaint keeps coming back:
“After converting, the spaces between words disappeared and I’m getting things like ‘mgoing’ and ‘corruptyour’.”
“My scanned PDF converted into pure gibberish — completely unreadable.”
“After converting a Chinese PDF, I got piles of characters I don’t recognize.”
These all point to the same stage — OCR (Optical Character Recognition). But even if your PDF isn’t a scan, the text-extraction step can still go wrong. Let’s break down what’s actually happening.
Two Kinds of PDF, Two Kinds of Problem
Text-based PDFs
PDFs exported from software like Word or LaTeX do contain text information. But the way that text is stored is far less friendly than you’d think.
A PDF stores text as a series of “place this character at coordinate (x, y)” instructions. The crucial problem: a PDF doesn’t necessarily store the spaces between words. The gap between two words might only be implied by a coordinate offset — the letter “m” sits at coordinate 120, the letter “g” at 135, and the space in between is hinted at by a 15-pixel gap.
A traditional converter has to judge for itself: “There’s a 15-pixel gap between 120 and 135 — at 10pt font, is that a space, or just normal letter spacing?”
It gets this wrong often. Too strict, and normal letter spacing gets read as spaces, splitting words apart. Too loose, and the spaces between words get swallowed, turning “I’m going” into “I’mgoing”.
Chinese PDFs add an extra layer of difficulty: Chinese itself doesn’t use spaces between words, but mixed Chinese-English text needs spaces inserted correctly where the language switches. Many converters crash and burn right here.
Scanned PDFs
Scanned PDFs are trickier — they’re essentially a stack of images with no text information in the file at all. OCR has to “read the letters.”
A traditional OCR engine (like Tesseract) works like this:
- Binarize the image (convert to pure black and white)
- Segment out the outline of each character
- Match against a template library of character shapes
- Output the recognition result
This process has a few hard weaknesses:
- Character-level recognition: each character is recognized in isolation, with no understanding of context. So “rn” and “m” get confused constantly (they look too alike), and “cl” vs “d” is another common mix-up pair.
- Layout sensitivity: if the scan is slightly skewed, the paper is wrinkled, or print quality is poor, accuracy plummets.
- Layout blindness: traditional OCR doesn’t understand page layout. It doesn’t know whether a block of text is body copy or a page number, and it doesn’t know how to read a two-column layout.
A user on the MobileRead forum ran a real test: a 129-page multi-column PDF came out at only ~60% accuracy with a traditional converter, the multi-column format was completely lost, and footnotes were folded into the body.
What Does AI OCR Do Differently?
pdf2epub.ai uses Google Gemini Flash — a multimodal large language model. The way it processes PDF pages is fundamentally different from traditional OCR.
Difference 1: Visual comprehension vs character matching
Traditional OCR does pattern matching: this pixel blob looks like the letter “a,” that one looks like “b.”
Gemini does reading comprehension: it “looks” at the whole page and understands what the text is saying. When it sees “I’m going,” even if some characters are blurry, it knows this should be “I’m going” and not “I’mgoing” — because it understands English grammar and meaning.
This means:
- The space problem basically disappears. The AI understands word boundaries; it doesn’t need to guess from pixel gaps.
- Look-alike character confusion drops sharply. The “rn” vs “m” problem gets resolved by context. For “mo__ing,” traditional OCR might output “moroing,” while the AI knows it should be “morning.”
- Mixed Chinese-English is handled automatically. The AI understands the language switch and inserts the right spaces in the right places.
Difference 2: Page-level understanding vs character-level processing
Traditional OCR segments characters first, then recognizes them one by one. What it sees is a series of isolated character outlines.
Gemini sees the whole page. It understands the roles and relationships of every element on the page at once:
This page has:
- A heading ("Chapter 3: Foundations of Quantum Mechanics")
- Two columns of body text
- A footnote at the bottom of the left column
- A page number "42" in the top right
- A mathematical formula at the bottom
- A diagram of an experiment in the middleAll of these judgments are made simultaneously and corroborate each other. The AI won’t mistake a page number for part of the body, because it understands the page number’s visual position and function.
Difference 3: Content-aware post-processing
After traditional OCR outputs raw text, it usually only does a simple spell-check.
Our pipeline adds a layer of AI-driven merging and optimization after OCR:
- Cross-page stitching: if a sentence is cut off at the bottom of page 40 and continues at the top of page 41, the AI joins them correctly.
- Footnote marker matching: the AI detects footnote markers in the body (like ¹²³) and correctly pairs them with the footnote content at the bottom of the page.
- Formula normalization: math symbols in the PDF (like x², ∑, ∫) are converted to standard LaTeX format to ensure they render correctly in the EPUB.
My PDF Is a Scan — Can It Be Converted?
Yes.
In fact, scanned PDFs are the scenario where AI OCR has the biggest advantage over traditional OCR. Because:
- A scanned PDF has no internal text structure to parse, so the traditional approach and the AI approach start from the same place — an image.
- AI’s visual comprehension is noticeably more robust on blurry, skewed, low-resolution images.
- AI can use context to infer blurry characters, which character matching simply cannot do.
To be honest, though: if your scan is extremely low quality (say, a fax that’s been photocopied several generations over), no OCR will produce a perfect result. AI will do considerably better than traditional OCR, but it’s not magic.
Performance Across Document Types
| Document type | Traditional OCR | AI OCR |
|---|---|---|
| Plain-text novel | Good (90%+) | Excellent (99%+) |
| Academic paper (two-column) | Poor (60-70%) | Good (95%+) |
| Technical manual (with code) | Poor (code formatting lost) | Good (code blocks preserved) |
| Math textbook (with formulas) | Poor (formulas garbled) | Good (converted to LaTeX) |
| Mixed Chinese-English | Poor (space problems) | Good (handled automatically) |
| Low-quality scan | Poor | Medium to good |
Try It
Rather than take my word for it, test it yourself: pdf2epub.ai gives you free credits on signup — take the PDF that’s been giving you the most trouble and convert it. You can preview a few pages first; if you’re not happy with the quality, you’re not charged.