Converting audio to text for free works along three routes. They differ in allowance, privacy, convenience and, depending on your material, in quality too. First, the free tiers of the online services: Descript processes 60 minutes of media per month but watermarks exports on the free tier, Notta gives you 120 transcription minutes a month and caps a single conversation at three minutes, Happy Scribe offers ten minutes of AI transcription to try, and f4x, which processes on its own servers in Germany, gives you 15 minutes with no subscription and no payment details. Second, the tools you may already pay for: Word transcribes uploaded audio files up to 300 minutes a month, but only with a Microsoft 365 subscription, and voice typing in Google Docs is free yet takes no file, only a live microphone. Third, your own machine: Whisper is MIT-licensed, noScribe is GPL-3.0, both cost nothing, have no minute allowance and need no account. That third route is the only one where the recording never leaves your device. You still pay for it, in setup time and compute time: noScribe budgets roughly one to three hours of processing per hour of interview.
I transcribe my own calls locally, and the reason isn't thrift, it's the question of who gets to see the recording. For a first client conversation the answer is: nobody but me. What runs on my own hardware and what I deliberately send to the cloud is on the AI page, and in more detail in Local AI: What Actually Works in 2026 and AI and Privacy: What the AI Gets to See.
Three routes, three limits
Free means something different on each of the three routes. This overview says what you run into first.
The column that decides it is usually the last one. For a voice note to yourself it doesn't matter. For a recorded client conversation it decides whether a data processing agreement is due.
The free tiers and what their allowances cover
All figures taken from the providers' own pages, retrieved 20 September 2026.
The cap is rarely just a minute count, as the right-hand column shows. As a trial run these tiers are still enough: ten minutes is plenty to see how a service copes with your microphone, your room and your vocabulary.
If you regularly need more, eight tools with processing location, speaker separation, formats and prices sit side by side in Transcription Software Compared.
Built-in: what Office and your browser already do
Word transcribes uploaded files. The feature is called Transcribe, sits in Word for the web and in Word for Windows, and accepts WAV, MP4, M4A and MP3, with German and English on the language list. Microsoft states 300 minutes of uploaded audio per month for users with a Microsoft 365 subscription, and up to 30,000 minutes with a Copilot licence. Two caveats come with it: without a subscription or Copilot licence the feature isn't there, and it needs an internet connection, so the processing does not happen on your machine. Anyone already paying for Microsoft 365 has the shortest path from file to text in a document here.
Voice typing in Google Docs costs nothing but takes no file. It runs in Chrome, Edge and Safari, needs a microphone and covers more than 120 languages and dialects. What it can't do is stated in Google's own help: it dictates live, and an upload of an existing recording isn't provided for. Useful for speaking your notes right after a meeting, useless for yesterday's recording; that one is more likely sitting on your phone, and what its built-in tools do with voice recordings is a topic of its own.
Your conferencing tool is probably already taking notes. Teams, Zoom and Meet bring their own transcription, enabled depending on licence, which saves you the detour via a file. Where those switches live and what the automation then produces as minutes is in Transcribing Calls in Microsoft Teams and Automatic Meeting Minutes with AI.
The route across your own machine
This route is the only one with two properties at once: permanently free with no allowance, and the recording never leaves the house.
Whisper directly. OpenAI's model, code and weights included, is under the MIT licence. You install it with pip install -U openai-whisper, and a run looks like whisper recording.wav --language German. There are six sizes, and the choice decides speed and memory.
noScribe, if you'd rather avoid the command line. The application is open source under GPL-3.0, costs nothing, runs on Windows, macOS and Linux, and brings speaker recognition, which plain Whisper does not. It covers around 60 languages, and output comes as HTML for opening in Word, as plain text or as WebVTT. The developer states two numbers openly: about three gigabytes of installed size, because the models ship with it, and about one to three hours of compute per hour of interview, longer on older machines, considerably faster with an NVIDIA card.
In the browser, with no installation at all. The Whisper Web project implements Whisper through Transformers.js and describes itself as speech recognition directly in your browser, with experimental WebGPU acceleration on a separate branch, also under the MIT licence. It is the shortest way to try the local variant before installing anything. Don't equate it with an installed offline application, though: the page itself and the model files come off the network, and what other connections a particular instance opens only shows once you look at the network traffic in your browser.
No account is involved in any of these three variants. If you are searching for a way that doesn't ask you to sign up, this is where you land: there is no provider to sign up with.
Setup is the real effort on this route, and it is a one-off. If you'd rather not work through it alone, bring the question to the community: there are people there who already run the same setup on their own machines, and on a call it is quickly sorted out on a shared screen.
What free actually means in the arithmetic
The local route isn't the better one, it's the other one. It trades money for time, and whether that trade works out depends on your material; the counter-calculation of when having a transcript made pays off is in its own post.
On top of that comes a point that applies to every Whisper-based tool, including the services that run Whisper under the hood: the project's model card states explicitly that the output may contain text that was never spoken, that the model is prone to repetition, and that it performs unevenly across different accents and dialects. In practice that means a transcript doesn't look patchy where the audio was unclear, it looks plausibly wrong. Proofreading is therefore part of the job, and those minutes belong in the arithmetic whichever of the three routes produced the text.
A paragraph on the law before you upload anything
Section 201 of the German Criminal Code makes it an offence to record another person's non-publicly spoken word without authorisation. Secret recording is out, and in everyday business the safe rule is prior express consent from everyone else involved. Processing it afterwards through a service is a separate question and needs its own basis under the GDPR. Whisper's own model card explicitly advises against using the model on recordings made without the consent of those recorded. I'm a developer, not a lawyer, and this is not legal advice. The long version with the statutes and the consent routes is in Transcribing Calls: What German Law Says.
Once the text exists, the work begins
The search for a free tool usually ends at the same place: you now have text. After thirty conversations you have thirty files, and the real question is no longer what the conversion cost, but who works through them.
That role is filled in my business by Gustav, my AI employee for call analysis. His sequence matches the third route above, and that is no coincidence: speech recognition runs through local Whisper, and the raw transcript never leaves the machine. Then comes the fixed anonymisation pass that turns names into roles, companies into industries, amounts into orders of magnitude and places into regions, and strips sensitive passages. Only the anonymised extracts go any further.
His output, accordingly, isn't the transcript but a pattern register: the recurring questions from many conversations, each with your own clearest answer so far, from which a playbook, FAQ answers, content raw material and an objection library follow. He works in stages of ten to twenty calls with a sample review, and he distils what you said yourself instead of inventing advice. Which roles carry the weight in a coaching practice is in AI for Coaches: Practice Over Hype.
Common questions
How do I convert audio to text for free?
Fastest through a service's free tier: f4x gives 15 minutes with no payment details, Notta 120 minutes a month. For an ongoing, uncapped option, use Whisper or noScribe on your own machine. Both are open source, cost nothing and need no account.
Does it work without signing up?
Yes, on the local route. Whisper and noScribe are programs on your own computer, and there is no provider to register with. Whisper Web brings the same model family into the browser, though it loads the application and the model off the network. The free tiers of the online services generally do require an account, though f4x at least skips payment details for the first 15 minutes.
Can Word convert audio to text?
Yes, through the Transcribe feature in Word for the web and Word for Windows. It takes WAV, MP4, M4A and MP3 and handles German as well as English. It requires a Microsoft 365 subscription or a Copilot licence, so it is free only if you already pay for that.
How long does the free route on my own machine take?
That depends on hardware and model size. noScribe budgets about one to three hours of processing per hour of interview, longer on older machines and considerably faster with an NVIDIA card. With Whisper directly you steer it through model size: tiny runs roughly ten times as fast as large per the project's figures, turbo roughly eight times.
Are free tools less accurate than paid ones?
Not necessarily, because a related model family, often Whisper, sits under many services and you can run it locally yourself. Model size, pre- and post-processing and speaker separation still differ, and so, depending on your material, does the result. I found no independent comparative measurement of these services for German, so your own ten-minute test with typical material tells you more than anyone else's percentage.
Is the free option also the more privacy-friendly one?
Locally the number of recipients drops: if no third party really processes anything, the data processing agreement falls away. Check cloud backup, synced folders and your tool's optional cloud features for that. Access control, legal basis and deletion periods stay your job even locally. For the free tiers of online services the same applies as for their paid tiers: the file goes to a provider. Price says nothing about this, processing location does.
Where to go from here
Take ten minutes of a typical recording of your own, not a clean studio take, and run it through exactly two routes: a free tier of your choice, and once locally, most easily through noScribe. Don't measure the hit rate, measure the time you need to correct it until you'd show the text to someone. After that you'll know whether price is the interesting number in your case, or correction time.
If it turns out the conversion isn't the bottleneck but the working-through afterwards is, Gustav is available as a finished package in my community: with the anonymisation pass, consent templates and a fictional demo transcript to practise on, in the premium tier at 45 dollars a month, as of September 2026. The templates are samples without warranty and don't replace a lawyer's look at your case. Everything about it is at Claude Practitioners.
