About ParallelTexts

ParallelTexts takes two versions of the same book — one in your native language, one in the language you are learning — and automatically aligns their sentences so you can read them side by side, sentence by sentence.

The output is a parallel corpus: a structured list of sentence pairs where each source sentence is matched to its translation. You can read it directly in the browser as a paginated ebook-style view, or export it for use in other tools.

• Language learners who want to read a novel in a foreign language with a trusted translation always one tap away.
• Linguists and translators who need to build small parallel corpora quickly, without setting up a pipeline.
• Researchers looking for a lightweight, reproducible alignment tool that runs entirely on their own machine.

The tool is designed to be approachable to anyone — no command-line knowledge required.

Under the hood, ParallelTexts runs a three-stage ML pipeline — entirely inside your browser using WebAssembly:

1. 1. Sentence splitting — Each book is split into individual sentences.
2. 2. Multilingual embedding — Every sentence is converted to a vector (a list of numbers that captures its meaning) using a multilingual AI model. Because the model understands 50+ languages at once, semantically equivalent sentences in different languages end up with similar vectors.
3. 3. Needleman–Wunsch alignment — A dynamic-programming algorithm (the same family used in DNA sequence alignment) finds the globally optimal pairing of source and target sentences based on their vector similarity.

The result is a list of aligned pairs tagged as 1:1 matches, or source/target gaps where no counterpart was found.

The AI model (~500 MB–1.1 GB depending on which you choose) is downloaded once from our servers and then cached permanently in your browser. Subsequent alignments reuse the cached version — no re-download needed.

Alignment is compute-intensive. A dedicated GPU is highly recommended — it can be 10–20× faster than running on CPU alone. Browsers with WebGPU support (Chrome 113+, Edge) will use your GPU automatically; others fall back to WebAssembly on the CPU.

You can manage which models are cached from the Settings page.

• EPUB — full support including cover art and chapter structure.
• PDF — text extraction via PDF.js (quality depends on the PDF).
• TXT — plain text files.

Japanese EPUBs are preprocessed to strip furigana before alignment so that phonetic readings do not interfere with the embeddings.

After alignment completes you get two reading modes:

• Popover view — Read the source text as a paginated ebook. Tap any sentence to see its aligned translation in a popover.
• Side-by-side view — Both languages displayed in parallel columns, every sentence pair on the same row.

You can also export any alignment in two formats:

• EPUB — a fully readable parallel ebook. Source sentences are shown by default; click any entry to reveal its translation. Images from the original books are preserved. The EPUB also embeds a hidden manifest so it can be imported back into ParallelTexts with the full alignment intact (see below).
• TSV — a tab-separated file with one sentence pair per row and an optional confidence score in the third column. Gap rows are preserved with an empty cell on the missing side. Any 2- or 3-column TSV works on import — not just files from ParallelTexts.

Both formats can be imported back into ParallelTexts from the Alignments page. Importing a ParallelTexts-exported EPUB restores the alignment exactly — pairs, images, and metadata — with no need to re-run the AI model. Importing the same EPUB as a book (on the Books page) shows only the source text, making it usable as a standalone foreign-language book.