Turning Audio and Video Into Text You Can Actually Use
Recording quality decides transcript quality more than the model does. How to prepare a file, what accuracy to expect, and why interviews and meetings should not be uploaded to a stranger's server.
Updated August 24, 20267 min read
Speech recognition has quietly become good enough that the model is rarely the limiting factor any more. What limits a transcript now is the recording: how far the microphone was from the speaker, how much the room echoed, whether two people talked at once. Feed a modern recogniser a clean recording and you get something close to publishable. Feed it a phone on a table in a meeting room and no amount of software rescues it.
What accuracy actually depends on
Transcription accuracy is usually quoted as a word error rate — the proportion of words inserted, deleted or mistaken. A good result on clean speech is under 5%, which is roughly one error every two sentences and easy to correct by reading. Above about 20% the transcript takes longer to fix than to retype, and the useful thing to know is which conditions push you across that line.
| Condition | Effect | What to do about it |
|---|---|---|
| Microphone close to the speaker | Best case, under 5% | Nothing — this is the goal |
| Phone on a table, two metres away | Noticeably worse | Move it closer next time; it is the single biggest lever |
| Echoey room, hard surfaces | Severe — reverb smears word boundaries | Soft furnishings, or record closer |
| Constant background hum | Moderate | Noise reduction helps a lot here |
| Two people talking at once | Both lost | Unfixable after the fact |
| Strong accent or fast speech | Somewhat worse | Expect more names and technical terms to be wrong |
| Music under speech | Bad | Unfixable in a mixed recording |
Notice how many of those rows are decisions made before the recording exists. Transcription is one of the few tasks where five seconds of preparation beats any amount of subsequent processing, and the preparation is almost always just: put the microphone nearer the mouth.
Preparing a file so it transcribes well
Get the audio out of the video
If your source is a screen recording or an interview shot on camera, extract the audio track first. Pick the format the tool marks as a lossless copy — usually M4A for an MP4 — and the track is copied out without re-encoding; working with a 30 MB audio file instead of a 2 GB video makes everything afterwards quicker. The free plan takes videos up to 300 MB and extracts up to an hour of audio; Pro removes those caps for MP4, MOV, WebM and MKV.
Clean up steady background noise
Air conditioning, computer fans and traffic rumble are constant, which makes them exactly the kind of noise that noise reduction removes well. Use a light setting — aggressive reduction eats consonants, and consonants are what the recogniser relies on.
Trim the parts with no speech
Ten minutes of setup chatter before the interview starts is ten minutes of processing you do not need. The silence remover collapses long gaps, and cutting the head and tail by hand takes seconds.
Split anything very long
Long recordings work on Pro (the free plan covers 10 minutes per file), but a failure two thirds of the way through costs you everything. Splitting a long file into fifteen or twenty minute sections makes the job resumable and easier to proofread against the audio.
Transcribe, then read it against the audio once
Do not proofread the text on its own. Play the recording at 1.5× while reading the transcript — errors that look plausible on the page are obvious when you hear what was really said.
What the model will get wrong
The errors modern recognition makes are not random — they are confident, grammatical and plausible, which makes them harder to spot than the garbled output of older systems. Knowing the categories tells you where to look.
- Proper nouns. Names of people, companies and products are the most frequent errors by a wide margin, because they are rare words the model has little reason to prefer. Build a list of the names in the recording and search for them afterwards.
- Homophones in context. *Their*, *there*, *they're*; *principal*, *principle*; *effect*, *affect*. All grammatical, all silently wrong.
- Numbers and units. Figures spoken quickly are often transcribed as words, or the wrong figure entirely. Anything financial or medical needs checking against the audio directly.
- Speaker attribution. Where a transcript separates speakers, the boundaries drift after interruptions and can swap two people for a whole passage.
- Invented fluency. Recognisers trained on written text tend to tidy speech — dropping false starts and repairs. Usually welcome, occasionally a problem when the exact words matter.
- Hallucinated phrases in silence. Long silent or musical passages can produce plausible sentences that were never spoken. Trim them out beforehand.
Why this should not be an upload
Think about what typically gets transcribed. Interviews with sources. Medical consultations. Legal calls. Board meetings. Therapy sessions. Research conversations under an ethics approval that specifies where the data may be stored. The category of audio people want transcribed overlaps almost exactly with the category of audio that should not be handed to a third party.
The standard online transcription service uploads your recording, processes it on shared infrastructure and returns text. The recording sits in their storage under a retention policy you did not negotiate, may be used to improve their models depending on what you agreed to, and is subject to whatever legal process applies where their servers are. For a great deal of professional recording, that alone makes the tool unusable regardless of accuracy.
When you transcribe a file, the speech to text tool here runs the recognition model on your own machine: the recording is processed on your device, not our servers, and the transcript stays with you. That also removes the upload wait entirely — a two-hour recording starts processing immediately rather than after twenty minutes of transfer (free: up to 10 minutes and 50 MB per file; Pro: as long as your device can hold) — and once the model has downloaded the first time, file transcription keeps working without a connection. Live microphone dictation works differently: it uses your device's speech service, which may run in the cloud, and the tool labels each session on-device or cloud.
File transcription that runs on your device, not our servers — no upload wait, no server copy, no retention policy to read.
Transcribe a recordingTurning a raw transcript into something useful
Raw transcription output is a wall of text: no paragraphs, inconsistent punctuation, and every *um* faithfully preserved. What you do next depends on what the transcript is for, and the four common destinations want quite different things.
| Purpose | What to keep | What to cut |
|---|---|---|
| Subtitles | Everything, with timings | Nothing — accuracy to the spoken word matters |
| Quotation in an article | Exact words, checked against audio | Nothing, but mark any ellipsis honestly |
| Meeting notes | Decisions, actions, numbers | Almost all of it — this is a summary task |
| Blog post from a talk | Structure and argument | Fillers, repetition, spoken syntax |
| Searchable archive | The full text as-is | Nothing; imperfect text is still searchable |
For the tidying itself, the text trimmer removes stray whitespace and blank lines in bulk, and the word counter tells you what you are working with — an hour of conversational speech is typically eight to ten thousand words, which is a useful thing to know before promising somebody a two-page summary.
One habit worth adopting from professional transcription: keep the raw output as a separate file and edit a copy. Once fillers and false starts are removed, they cannot be recovered without going back to the audio, and the moment you need the exact original wording is invariably after you have cleaned it.
Frequently asked questions
How accurate is automatic transcription?
On a clean recording with the microphone close to the speaker, modern recognition typically errs on fewer than one word in twenty, which is quick to correct by reading. Distance from the microphone, room echo and overlapping speakers each degrade that sharply, and no software recovers what a bad recording never captured clearly.
Can I transcribe a video directly?
Yes, though it is faster to extract the audio track first. Extraction is quick — and lossless when you keep the track's own format — and it turns a large video into a modest audio file, which makes cleaning, splitting and transcribing it considerably faster afterwards.
Is it safe to transcribe confidential recordings online?
Only if the processing happens on your own device. A service that uploads the audio has a copy of it, subject to their retention policy and jurisdiction, which rules out most professional recordings — interviews, consultations, legal calls. Transcription that runs locally never sends the file anywhere, so there is nothing to retain.
Why does the transcript get names wrong?
Proper nouns are the hardest category for any recogniser, because they are rare and the model has no contextual reason to prefer an unusual spelling over a common word that sounds similar. The efficient fix is to list the names appearing in the recording and search for near-misses afterwards, rather than reading for them.