All posts

EngineeringSeptember 16, 2026 · 5 min read

Transcribing a 10-hour recording without falling over

Most transcription services quietly cap out around two or three hours. Here's the checkpointed pipeline that lets Turnlogue take a ten-hour file and return a verified, topic-structured summary.

Upload limits are usually presented as storage constraints. They're not. They're compute-time constraints: the moment your processing pipeline has to finish inside a single function timeout, your real product limit is however many minutes of audio that timeout can absorb. Most services cap out somewhere around two or three hours and let you discover it with a failed upload.

Step one: get the file there at all

A ten-hour WAV can weigh gigabytes. No browser uploads that in one request. Turnlogue uploads in resumable chunks straight to object storage — a dropped connection at hour nine costs you the last chunk, not the file. When the last piece lands, the transcription job starts against the stored file directly.

Step two: never trust a single run

The transcription itself runs as a durable, checkpointed job. The pipeline is split into phases — map the transcript in chunks, reduce each chunk into structured findings, verify coverage, persist, translate — and every completed phase is recorded before the next begins. If a run hits its time budget mid-way, it stops, saves its position, and the next invocation picks up exactly where it stopped. No phase re-reads work its predecessor already finished.

This is the difference between "supports long files" as a marketing claim and as an architecture. A 40-minute file and a 10-hour file exercise the same code path — the long one just iterates more.

Step three: summaries that survive length

Summarizing a ten-hour transcript by asking a model to "summarize this" fails twice: once because the text doesn't fit, and again because even if it did, a single-pass summary of ten hours is mush. Turnlogue's pipeline maps over the transcript in chunks — extracting topics, decisions, and action items per segment — then reduces those findings into a coherent whole, and finally runs a verification pass that checks the summary against the transcript to catch anything dropped between chunks.

Step four: translations in the background, not the way

Translation of every turn into your chosen languages starts with the summary and fills through the transcript progressively — the record is readable before the last language finishes landing, and nothing blocks on it.

The user-visible promise is simple: the length of your recording is a cost question, never a capability question.

Long recordings are where meeting notes go to die — nobody reads ten hours of raw audio time. A verified, topic-structured record is what makes them useful. Try it with your longest file.

Put it to work on your own conversations

Live sessions, 10-hour uploads, translations into 50+ languages. Free plan included.

Start free