Overtone: open-source captions that re-cut per screen shape

Overtone is an open-source captioning tool that Show HN user rvanderlinden posted on 29 October 2026; it reached 42 points. According to its README, it stores each spoken word once, with its time and speaker, in a .overtone.json file, then cuts captions separately for each picture shape: 16:9 and 9:16 by default. A command-line tool builds the file from subtitles, audio or video, and a video.js plugin shows the cut that fits the player. It is MIT-licensed, made by Lont.

What it runs on

The README describes several engines hearing the audio. By default, gpt-transcribe supplies the words and Azure AI Speech supplies word times and speakers; ElevenLabs Scribe can be added as a second opinion. Where they disagree, a language model re-listens to the snippet and picks the words.

An LLM editor then corrects misheard names and terms using your context and glossary, labels speakers and marks key words. The README says the model never sets timestamps, code checks only the structure of its changes, and edits it is unsure about go on a review list. A `--local` mode runs offline with whisper.cpp and Ollama, and a `--dry-run` flag shows where audio goes and the estimated cost before anything is sent.

What the author measured

All figures are the author's, from a small public benchmark. On six clips in four languages (20 minutes, three runs, 23 October 2026), the README reports a pooled word error rate of 5.3% on the five clips where every engine returned text, against 6.2% for gpt-transcribe alone and 7.7% for Azure Speech alone. It reports $0.023 to $0.025 per audio minute at list prices, or $0.029 with ElevenLabs Scribe.

On line breaks across three English clips, the author reports 0.3% of Overtone's breaks end on a word such as "the" or "of", against 28.8% when lines are simply filled to a length limit.

The failure worth copying the lesson from

The benchmark page says gpt-transcribe returned empty text for a quiet meeting recording in every run. Overtone retried with the language the other engines had heard and scored 32.8% word error on that clip, against 42.9% for Azure Speech alone. The author's follow-up test found that setting the language explicitly made gpt-transcribe return about 700 words. If you run transcription in production, alarm on empty output and pass the language.

Limits and pushback

The README says full line-cutting rules exist only for English and Dutch, there is no translation or sound tags such as [music], and media is limited to two hours. The benchmark is small, and the author notes the numbers are scored with their own normaliser, so they compare engines with each other, not with published leaderboards.

The thread had two replies. One commenter said modern subtitles are full of misheard words and asked for translation of on-screen foreign text, which the README lists as not supported. Another said the cost of speech-to-text is real and hard to explain to end users, which is the argument for the tool's cost estimate.