Claude coded it: a 60-hour audiobook synced to its text in 20 minutes
For syncing epub and audiobook files, they had Claude implement forced alignment using a CTC transcription model. It aligns a 60 hour book's transcript in 20 minutes, even with dirty source data.
Done withClaudeSpotted· by HN user nodja· Oct 22, 2026
Result Runs on their own GPU with a 0.6B model
“This method aligns the transcript of a 60 hour book in 20 minutes on my GPU using Parakeet CTC 0.6B”
Spotted: found in the wild. Details are as reported by the source above and have not been independently verified.
What they did
Wanting really accurate timestamps for syncing epub and audiobook files, they asked Claude to implement forced alignment using a CTC transcription model, matching the logprobs of a character based transcription model to find the best overlap between audio and text.
How it went
The technique generates sync files that are really good even when the data is dirty. It aligns the transcript of a 60 hour book in 20 minutes on their GPU, using the English only Parakeet CTC 0.6B model, with similarly sized models available in other languages.
Try it yourself with Claude
Implement forced alignment to sync [a transcript/ebook text] with [an audio recording] using a CTC transcription model: use its character level logprobs to find the best matching overlap between the audio and text, so I get accurate timestamps even when the source text is messy.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.
A browser game where you answer with a power of ten leans on Claude for code and content: fact checking answers, writing explanations, and inventing new questions.
A new live-coding tool, Strudel, and no experience using it. Claude wrote the starting code, including sample filenames that turned out not to exist, then restructured it for live performance once it was tweaked by hand.
They wanted a blog article turned into a video without losing control of the text. Claude built a recording app to screencast the article's diagrams and add titles. It worked, but it still took them many hours.
They gave Claude the video, its audio and ffmpeg, and asked for a code-only recreation. One pass, under an hour, and only the end logo differed. The writer calls it "just an experiment".
She has no tech skills and very specific ideas. Built with Claude over a week of evenings: drop a folder with text and images, preview it at home, click to publish. She has exactly the site she wants.
She invented the rules, the name and the board: words climbing from 3 to 7 letters in a pizza-slice shape. Claude turned her design into code, and it became a real daily puzzle over a weekend.