How to convert audio to text: the three routes and what each costs
How to convert audio to text three ways: type it yourself, upload to a cloud service, or run speech recognition in your browser. Costs, speed and privacy.
There are three ways to convert audio to text, and only three. You can type it out yourself, upload the file to a cloud transcription service, or run a speech recognition model on your own computer, which now includes running it inside a browser tab. They differ in what they cost in money, in how much of your time they take, and in where your recording ends up. This page lays out all three so you can pick one, then sends you to the specific guide for whatever file you are holding.
The three routes to convert audio to text
If you just want to turn a voice recording into text once and move on, the decision takes about a minute. What separates the three is money, time and who else gets a copy of the audio.
Route one is manual: you open the file in a player and type what you hear. Route two is a cloud service: you upload the file, a model runs on a server, and text comes back. Route three is local: the model runs on your own hardware, which used to mean a Python install and now can mean opening a page in Chrome.
| Route | Money | Time for one hour of audio | Where the audio goes | Suits |
|---|---|---|---|---|
| Type it yourself | Nothing | Several hours, depending on the audio | Nowhere, it stays on your disk | Short clips, messy audio, verbatim work |
| Cloud service | Free tiers with caps, or a monthly seat, or per minute | Upload, then wait for the provider | Uploaded to the provider’s servers | Work a local English base model cannot do |
| Local model | Nothing beyond electricity | Roughly the length of the recording, sometimes less | Nowhere, it stays on your disk | Confidential audio, long files, no account wanted |
Typing it yourself is slow, and sometimes still the right call
Manual transcription costs nothing and gives you complete control. It is the only route where you decide, sentence by sentence, whether to keep the false starts.
The honest way to estimate it is to measure rather than guess. Transcribe the first five minutes of your recording, note the clock, and multiply. Typing is not the slow part. Stopping, rewinding four seconds and playing the same phrase three times is.
Two things make it bearable: a player with a keyboard shortcut for a short rewind, so your hands never leave the keys, and playback slowed to around 70 per cent rather than constant stopping and starting.
Type it yourself when the recording is under a couple of minutes, when several people talk over each other, or when you need verbatim conventions that no automatic tool applies. If the audio is difficult because of how it was captured, the fix belongs at the recording stage, and nine ways to record audio that transcribes well covers the habits that matter.
A cloud service buys convenience with money and a copy of your file
Upload, wait, download. If you need speaker labels, summaries, a searchable archive, a calendar integration or a language other than English, look at the plans on this route, because a local English base model gives you none of them.
Published prices give you the shape of it. Otter’s Basic plan is listed as free, with 300 monthly transcription minutes, a 30 minute maximum conversation length and three lifetime audio or video file imports. Its Pro plan is listed at $16.99 per user per month, or $8.33 per user per month billed annually, with 1,200 monthly transcription minutes and ten monthly file imports. Rev lists paid subscription plans from $25.49 per seat per month, and human transcription at $1.99 a minute, which works out at $119.40 for an hour of audio. Those are the figures on the vendors’ own pricing pages on the date in the sources below.
Two things follow. Free tiers carry published caps, and the arithmetic behind the 30 minute cap shows where the line sits. And once you divide a monthly fee by the minutes you actually use, the price per hour moves a long way from the sticker, which is what the cost per hour comparison works through.
The other cost is not money. The recording leaves your machine and exists as a copy on infrastructure you do not control. That is often fine for a webinar you plan to publish anyway, and worth a deliberate decision when the audio is a client call, a patient, a job candidate or a source, because you are then handling someone else’s personal data. The checklist for uploading client calls lists the eight questions to settle first.
Local speech recognition runs the model on your own hardware
OpenAI released the Whisper speech recognition models with code and weights under the MIT licence, in sizes from tiny to large, with English-only variants of the four smallest. Anyone could suddenly run speech recognition without an API key.
WebGPU lets a web page use your graphics card for general calculation, so the model can be downloaded once, cached, and executed in the tab. The file is read from disk by the page, the same way an image editor reads a photo, and nothing is sent anywhere. What WebGPU is and why it matters explains the machinery, and the model sizes post covers what you give up by running a small model rather than a large one.
Speed depends entirely on your machine. In our test, 71 seconds of audio took 41 to 50 seconds including the first model load on a desktop with a graphics card. A thin laptop takes roughly as long as the recording itself. You do not have to watch it, but the tab has to stay open.
The trade is straightforward: you pay nothing and upload nothing, and in exchange you get a smaller model, one language and no speaker labels. Our transcriber takes this route, and the next section is where it stops.
Pick the route by the recording you are holding
Most people arrive with one specific file, and the route usually follows from where it came from.
- A Zoom recording: find the audio-only M4A in your Zoom folder first, as the Zoom guide describes.
- A Teams recording: it lands in OneDrive or SharePoint, and the Teams guide covers getting it out without admin rights.
- An iPhone voice memo: the M4A has to get to a desktop before anything else happens, which the voice memo guide walks through.
- Your own YouTube upload: download the MP4 from YouTube Studio, per the YouTube guide.
- A podcast episode: the show notes workflow goes from audio to publishable notes.
- A lecture recording: the student workflow covers the studying part, not just the transcript.
Whatever the source, the last step is choosing an export. Notes and quotes want plain text; captions want SRT or VTT. The three exports compared is a one-screen decision table if you are unsure.
Where our tool stops, and what to do about it
Being specific about the limits saves you a wasted download. Our transcriber needs desktop Chrome or Edge with a working WebGPU adapter, so Firefox, Safari and phones will not run it yet. It is English only. The first run downloads about 200 MB of model, once, and the browser caches it after that. It uses the base English model, chosen to run on ordinary hardware rather than for top accuracy. It produces one continuous block of text with no speaker labels, and speaker separation is a different task that the model does not perform. Nothing you open is ever uploaded, because there is no server to upload to.
Read the output before you rely on it. Names, jargon, strong accents and a bad headset are the usual failure points, and seven mistakes to fix in a transcript is the checklist we use on our own.
If those limits fit your file, convert your audio to text in the browser with no account and no upload. If they do not, one of the other two routes above will.
Frequently asked questions
What is the fastest way to convert audio to text? Either automatic route, by a wide margin, because typing out an hour of audio takes several hours. Between the two, a local model on a desktop with a graphics card ran at about 1.5x real time in our test, and there is no upload to sit through first.
Can I turn a voice recording into text for free? Yes, on two of the three routes. Typing costs only your time. Running a model locally costs nothing beyond electricity. Cloud free tiers exist and carry published caps on minutes, file length or the number of imports, like the ones quoted above.
Does the audio file format matter? Mostly no. MP3, WAV, M4A, AAC, FLAC and OGG all work in our tool, as do MP4, MOV, WEBM, MKV and AVI video files, from which only the audio track is read. What matters far more is what the microphone picked up. A clear recording in a compressed format beats a noisy one in a lossless format every time.
Will any of these give me a transcript with speaker names? Only the cloud route, and only on the plans that list it as a feature. Local speech recognition returns words and timings with no attribution. If you need labels from a local transcript, record each person to a separate track where the platform allows it, transcribe each track on its own, and merge them by timestamp.
Sources, checked 15 September 2026
- Otter.ai pricing, plans and limits: https://otter.ai/pricing
- Rev pricing, subscriptions and per-minute rates: https://www.rev.com/pricing
- OpenAI Whisper repository, licence and model sizes: https://github.com/openai/whisper
- Speed, model size and input formats: our own measurements on this site’s transcriber