From recording to usable notes: what to do with three hours of audio

Transcription got good and cheap at roughly the same time, which produced a new problem: people now have transcripts they never read. A three-hour recording becomes a thirty-thousand-word wall of text, and a thirty-thousand-word wall of text is not more useful than the recording. It is arguably less useful, because at least the recording had a timeline.

The interesting question stopped being can we transcribe this and became what shape should the output be.

Transcription accuracy is mostly an input problem

The engines are all roughly comparable now on clean audio. What separates a 97% transcript from an 85% one is almost never the model:

Overlapping speech is the single biggest killer. Two people talking at once produces a garbled merge, and no amount of model quality fixes it after the fact. If you control the recording, one microphone per speaker solves more accuracy problems than any other intervention.

Room acoustics matter more than microphone price. A cheap mic in a carpeted room beats an expensive one in a glass conference room with a hard table.

Domain vocabulary — product names, drug names, acronyms, people's names — will be wrong unless you supply them. Most serious tools accept a vocabulary hint list, and most people never use it. Ten minutes assembling one is the highest-leverage ten minutes in the whole pipeline.

Accents and code-switching. Systems trained predominantly on one accent degrade on others, and they degrade silently. If your speakers switch languages mid-sentence, test explicitly before trusting the output.

Pick the output shape before you record

This is where most of the value is, and it is a decision, not a setting.

Verbatim transcript. Every word, every "um." Correct for legal, medical, and research use where the exact utterance is the artifact. Almost useless for everything else.

Clean transcript. Filler removed, false starts smoothed. What most people actually mean when they say "transcript." Readable, still long.

Timestamped and speaker-labelled. The version that stays useful over time, because it stays navigable — you can jump to the moment rather than scrolling for it. If you expect to come back to this recording more than once, this is the format you want. Tools built around keeping a recording navigable rather than just converting it — Audien is one — are solving this shape of the problem specifically.

Structured summary. Decisions, action items, open questions, with timestamps back to the source. The most useful output for meetings, and the one where verification matters most: a summary that invents an action item nobody agreed to is worse than no summary.

Searchable index. For archives — dozens or hundreds of recordings. Here you don't want to read anything; you want to find the four minutes that matter.

Deciding this up front changes what you record, how you record it, and which tool you reach for.

The failure nobody checks for

Summarization models drop things they consider unimportant, and their sense of importance is not yours.

The specific pattern to watch: a decision that was reversed later in the meeting, where the summary captures the first version. A caveat attached to a commitment, where the summary keeps the commitment and drops the caveat. A dissent expressed briefly and then talked over, which disappears entirely.

None of these look like errors in the output. The summary reads clean. This is why "spot-check the summary against the timestamps" is a real step and not an optional one — and why timestamped output is worth the extra effort even when you plan to read only the summary.

When the recording isn't in your language

Two-stage — transcribe, then translate — beats direct speech-to-text-in-another-language for most content, because you keep the original transcript as an auditable artifact and can fix transcription errors before they propagate through translation.

The exception is live or near-live use, where the extra hop costs too much latency.

If the source material is a mix of recordings and documents — a research project, a due-diligence review, a literature survey — the documents need their own path rather than being forced through the audio pipeline. Digital files go to a document translator like 술술번역; scans and photographs need OCR first, which is what tools like Scanned.to and ScanRead.ai handle. Get everything into text, then treat it as one corpus.

A reasonable default

Record with one mic per speaker where you can. Supply a vocabulary list. Produce a timestamped, speaker-labelled transcript as the durable artifact, and generate summaries from that rather than instead of it.

The summary is what you read this week. The timestamped transcript is what makes the recording still worth something in six months, when you have forgotten it exists and need one specific sentence out of it.