GuidesMeetings
How to transcribe audio to text for free: methods and limits
Transcribe audio to text for free with Google Docs, your phone, Whisper or an online service: the steps, the limits, accuracy and GDPR.
Contents
- What free methods are there to transcribe audio to text?
- How do you transcribe an audio file with Word?
- How do you transcribe with Google Docs or your phone’s dictation?
- How do you transcribe audio with Whisper on your computer?
- What limits do free online services have?
- What lowers the accuracy of a transcript?
- Can you send a sensitive recording to a free service?
- Is a transcript enough to write the minutes?
- What about meetings that come back every week?
- Where do you start?
To transcribe audio to text for free, the right tool depends on where the speech is. A recording you have yet to make is transcribed on the phone itself, with Voice Memos on an iPhone 12 or later, Recorder on a Pixel or the Live Transcribe app on Android, while a file already recorded goes through Whisper, OpenAI’s open model that runs on your computer, or through the free tier of an online service, which stops at a few half-hour files. Word transcribes a file too, but only with a Microsoft 365 subscription, and Google Docs voice typing only hears the microphone.
The first page of Google for this query is made up almost entirely of vendors’ pages recommending their own tool, and of rankings in which every tool listed has a paid plan to sell. Three questions separate the methods faster than any ranking, though: how long the recording is, whether you need to know who said what, and how sensitive its content is, because a job interview or a client meeting does not get uploaded to the first site that comes along.
What free methods are there to transcribe audio to text?
The table brings together the methods you can use without paying, or without paying more than a subscription the company often has already, with what each one accepts, its limit, how it handles speakers and where the audio goes. The limits come from the vendors’ help and pricing pages, read on 1 November 2026, and they change more often than the features themselves.
| Method | What it transcribes | Limit without paying | Speakers | Where the audio goes |
|---|---|---|---|---|
| Word for the web, Transcribe | a .wav, .mp4, .m4a or .mp3 file, or a recording made in Word | no transcription without a subscription; 300 minutes of files a month with Microsoft 365 | “Speaker 1”, “Speaker 2”, renameable | transcribed online, file kept in OneDrive |
| Google Docs, voice typing | whatever the microphone picks up, live | free with a Google account, in Chrome, Edge or Safari | none | to the browser’s speech recognition service |
| Voice Memos (iPhone) | a recording made in the app | free on iPhone 12 or later, in some countries | not mentioned by Apple | not stated by Apple |
| Recorder (Pixel) | a recording made in the app, French included, in real time | free, on Pixel phones | numbered, in US English only | voices labelled on the phone; a new transcription may go to Google |
| Live Transcribe (Android) | whatever the microphone picks up, live | free | none | history kept three days at most; French offline on Android 12 and later |
| Whisper (computer) | an audio file ffmpeg can read | free and open (MIT licence), with Python and ffmpeg | none | stays on your computer |
| Free online tier | an uploaded file | TurboScribe: 3 files of 30 minutes a day; Otter: 300 minutes a month, 3 imports in total | depends on the service | the vendor’s servers |
The table brings out a limit shared by every method: no free method puts real names on the voices, and the best of them settle for numbered speakers that you then rename by hand. It also shows that two much-cited methods, Docs voice typing and Live Transcribe, do not transcribe a file but what the microphone hears, which makes them tools for live speech far more than for a recording that already exists.
How do you transcribe an audio file with Word?
Word transcribes an audio file in its web version, provided you are signed in with a Microsoft 365 subscription in Edge or Chrome, and up to 300 minutes of uploaded audio a month. For a company that already has it, this is often the simplest method, since the text lands straight in a document you can rework and share.
- Open a document in Word for the web, signed in to your Microsoft 365 account in Edge or Chrome.
- On the Home tab, open the Dictate menu and choose Transcribe, which opens the pane of the same name beside the document.
- Select Upload audio, then pick your .wav, .mp4, .m4a or .mp3 file in the file picker.
- Keep the pane open while it processes, which can take up to the length of the recording, then add the transcript to the document, whole or in excerpts.
The service separates the voices and labels them “Speaker 1”, “Speaker 2” and so on, and a label renamed once changes throughout the transcript, which takes a minute for a meeting of four. The audio file is saved to the Transcribed Files folder in your OneDrive, where you can delete it once you no longer need to keep it, and Microsoft says your audio is sent to it and used only to provide the service. Microsoft lists more than eighty languages and regional variants, French from France and from Canada among them, and its help page says the transcription button only appears with an active Microsoft 365 subscription: that is what separates Word from the methods that are genuinely free.
How do you transcribe with Google Docs or your phone’s dictation?
Google Docs and phones transcribe live speech for free, and they suit a recorded file poorly. They are the right tools for a face-to-face interview, a lecture or a voice memo, provided the transcription is started while people are speaking.
With Google Docs voice typing
Google Docs voice typing is free and accepts French as well as English, but it transcribes what the computer’s microphone hears. Open a document in Chrome, Edge or Safari, choose Tools then Voice typing, click the microphone and speak at a normal volume and pace; a second click on the microphone stops the dictation.
For an existing recording, the only way is to play it beside the microphone, and the result then depends on the speaker, the distance and the room: a vendor such as Otter writes itself that its transcription is more accurate when it reads a meeting’s audio stream than when it goes through the device’s microphone. Two further limits apply: voice commands, punctuation included, only work with both the account and the document language set to English, so in French the full stops and commas are added by hand, and Google states that the browser controls the speech recognition service, which means the browser decides how your voice is processed.
With Voice Memos on the iPhone
On an iPhone 12 or later, the Voice Memos app transcribes what it records for free, during the recording or after it, in English, Spanish, Portuguese, Italian, French, German, Japanese, Korean, and Simplified and Traditional Chinese, in the countries and regions where Apple offers the feature. While recording, swipe up from the top of the waveform and tap the transcription button to watch the text appear; afterwards, open the recording, tap the options button, then choose Copy Transcript to take all of it, or View Transcript to select only part.
The app’s search finds a word across all your transcripts, and selecting a word in the text moves playback to that point in the audio, which makes checking a doubtful passage quick. Apple mentions no separation of speakers, on the other hand, and the help page does not say where the transcription is computed.
With an Android phone
On Android, Google’s Live Transcribe app, preinstalled on Pixels and available on Google Play for other phones, transcribes live whatever the microphone hears, in French as well as English, and even offline on Android 12 and later once the language is downloaded. It is mostly for following a conversation or a lecture: the history is kept for three days at most before it is deleted, its settings let you export the stored transcriptions, and a list of custom words helps with the product or client names a general model does not know.
Pixel phones also have Recorder, which transcribes French and English in real time and keeps each recording with its text. It only separates speakers in US English, and Google says the feature that transcribes a recording again in another language may process the audio on its servers.
How do you transcribe audio with Whisper on your computer?
Whisper is a speech recognition model published by OpenAI, whose code and weights are released under the MIT licence: you install it once on your computer, and it then transcribes as many files as you like, with no account, no length limit and without the audio leaving the machine. It is the most robust free method for a long or confidential file, and the least welcoming for anyone who has never opened a terminal.
Installing it takes Python, versions 3.8 to 3.11 according to the repository, and the ffmpeg tool, which installs with brew install ffmpeg on a Mac and with choco install ffmpeg or scoop install ffmpeg on Windows. Whisper itself then installs with pip install -U openai-whisper, and a single command transcribes a file while stating its language:
whisper meeting.m4a --language French --model turbo
The turbo model, which the command uses by default, needs about 6 GB of video memory according to the repository’s table; on a more modest computer, the small model makes do with about 2 GB at the cost of some accuracy, since the model sizes trade speed against accuracy. Whisper does not separate speakers, which the repository never mentions, and the repository warns that its performance varies widely from one language to another.
Some online services run on this same model, and TurboScribe says so on its home page. Installing Whisper then amounts to doing yourself what the service does for you, with the one difference that matters for a sensitive recording: the file goes nowhere.
What limits do free online services have?
Online services transcribe a file in a few minutes with nothing to install, and their free tier is bounded by the number of files and their length, two limits designed to make the paid plan tempting. Two examples taken from their help and pricing pages on 1 November 2026 show the gap between what a home page promises and what you get without paying.
TurboScribe transcribes three files a day for free, each up to 30 minutes long, uploaded one at a time and handled at a lower priority than subscribers’ files, while its paid plan takes files of ten hours. Otter’s Basic plan gives 300 minutes a month and 30 minutes per transcription, but allows only three audio or video file imports per account, in total rather than per month, so its free plan is for recording live far more than for transcribing files. Otter transcribes in English, Spanish, French, German, Japanese or Simplified Chinese, one language at a time, except for French and English, which it can follow together in the same meeting.
What these plans do not print in large type is the real price of free: your file is sent to the vendor’s servers, which keep it under their own rules. For a voice memo or a lecture the question hardly arises; for an interview or a meeting with a client it is the main question, and it has its own section below.
What lowers the accuracy of a transcript?
Accuracy depends on the recording first, well before the tool: a voice close to the microphone, one person speaking at a time and a quiet room give an almost clean text with any method in the table, whereas a distant recording made in an open-plan office stays mediocre with the best of them. Otter itself names background noise, accents and the complexity of the conversation among the factors that make its accuracy vary, and Google advises holding the phone close to the person you want to transcribe.
Distance from the microphone. A phone placed in the middle of a table picks up the person sitting opposite well and the one speaking from the far end of the room much less so. For an interview, put it between you and the other person; for a video call, record the computer’s sound rather than what comes out of the speakers, and check which microphone input is selected, since Microsoft warns that the wrong one gives disappointing results.
Overlapping voices. When two people speak at once, tools keep one, mix the two or lose one, and speaker separation goes wrong mostly in those passages. That is also where minutes most need to know who said what, which is why those moments are worth checking against the audio.
The company’s own vocabulary. Client names, product names and internal acronyms are the words most often mistranscribed, and they are exactly the ones that matter in the final text. Live Transcribe and Otter let you add a list of custom words, which is worth doing for a meeting that keeps coming back with the same names.
Mixed languages. Pixel’s Recorder transcribes the most spoken language when a recording contains several, and Otter follows only one language at a time apart from the French and English pair. A meeting that switches between languages is therefore transcribed badly unless the tool was chosen with that in mind.
In every case, test the method on five minutes of your own audio before running an hour of recording, then check the transcript against the sound for proper names, figures and dates, which are the most costly errors and the ones a quick read lets through.
Can you send a sensitive recording to a free service?
A recording in which someone’s voice can be recognised is personal data, and sending it to a transcription service means handing that processing to a third party, which makes confidentiality the first question for a job interview, a client meeting or an internal meeting. Before uploading a file, the service’s terms tell you in a few minutes where it will be stored, how long it will stay there and whether it may be used to improve the vendor’s models.
When a company uses a service to process its clients’ or employees’ data, article 28 of the GDPR requires it to use only processors that provide sufficient guarantees, under a contract that sets out the subject, duration and purpose of the processing. If the servers are outside the European Union, the transfer must also be covered, for example by an adequacy decision or by the Commission’s standard contractual clauses, as the French data protection authority, the CNIL, points out. A free account opened by an employee with a personal address meets none of these conditions, and it is the rule we apply to every consumer assistant, set out in detail in our piece on what ChatGPT can do for a recruiter: what contains neither a client nor a candidate can go there, and the rest goes through a tool the company has chosen and contracted.
Two methods in the table avoid the question because the audio never leaves the device: Whisper installed on your computer, and Live Transcribe used offline with a downloaded language. If employees are already transcribing their meetings on free sites, it is above all a sign that no tool was provided to them, and banning it without offering one only moves the use out of sight.
There remains the question that comes before all the others, that of the people being recorded, who must know before the recording starts. The French Criminal Code punishes with one year’s imprisonment and a €45,000 fine recording, without their consent, words spoken privately or in confidence, with consent presumed only when the recording was made openly in front of the people concerned, who could have objected. The French Labour Code further forbids collecting information about an employee through a device they were not told about beforehand.
For recorded calls, the CNIL asks that each person be told the purpose, the recipients, their right to object and their right of access, and nothing justifies doing less for a phone placed on the table. So you announce the recording, say what it will be used for and who will receive the text, and stop it if anyone asks.
Is a transcript enough to write the minutes?
No, because a transcript gives back everything that was said, in order, whereas minutes keep only what is useful after the meeting: the decisions, the actions with an owner and a due date, and the points left open. An hour of transcribed meeting produces a document nobody will read in full, and sending it as it stands passes on to every recipient the sorting the writer did not do, which is the very definition of work that looks finished and is not.
The transcript remains excellent raw material, provided you read it with the five sections of a set of minutes in mind, which we set out in our guide to writing meeting minutes. A summary produced by an AI is reread against the same grid, because it readily forgets an owner or a date, and the time is really saved afterwards, when each action is chased on its due date. That is the argument of our analysis of meeting note-takers: the minutes stay inside the company, whereas the follow-up sent to a client commits whoever signs it.
What about meetings that come back every week?
For a recurring meeting, transcribing a file after the fact is the slowest method, since every week someone has to remember to record, retrieve the file, upload it, rename the speakers and then write the minutes. A tool that joins the meeting itself removes those steps, provided the participants know and whatever leaves the company is still reread by a person.
That is what Balt does, the AI employee of small and mid-sized businesses. Once Google Calendar or Outlook is connected, he joins the Zoom, Google Meet, Teams, Webex or GoTo meetings that have a video link, transcribes them with each speaker named, including a meeting held in two languages, which he transcribes in both, and sends you the minutes as a direct message in Teams, Slack or by email, with the summary, the decisions, the actions to take and the key quotes. Whatever then has to leave the company, the recap for the client as much as the reminder to a supplier, is put to you before it goes, under the rule we apply to everything an AI writes to the outside world.
For an audio file already recorded, the free methods in this article remain the right ones, and the details of what Balt does in meetings are on the meetings page.
Where do you start?
With the three questions from the beginning. If the recording has yet to be made, start your phone’s transcription and put it close to the person speaking; if it is a short file with nothing at stake, an online service’s free tier is enough, or Word if you already have Microsoft 365; if it is a long or confidential file, install Whisper once and for all. In all three cases, test on five minutes, check names and figures against the audio, and do not mistake the text you get for the minutes that still have to be written.
When the same meeting comes back every week, its minutes become a good task to hand to someone else, colleague or agent, because they can be checked at a glance: that is the test we recommend for choosing the first tasks to delegate.
Frequently asked questions
Is there a free AI that can transcribe audio to text?
Yes. Whisper, the speech recognition model published by OpenAI, is free and released under the MIT licence: it installs on a computer with Python and the ffmpeg tool, then transcribes files with no length limit and without sending the audio over the internet. Without installing anything, online services offer narrower free tiers, for example three 30-minute files a day at TurboScribe, or 300 minutes a month but only three file imports per account at Otter.
Can ChatGPT transcribe an audio file?
Yes, but not on the free plan: OpenAI keeps audio uploads for paid subscriptions, for files up to 512 MB in WAV, MP3, M4A, FLAC and a few other formats. OpenAI itself warns that transcripts may contain errors and that speaker identification may be unreliable. For an interview or a client meeting, a personal account is not a tool the company has chosen, and the GDPR question comes before the quality one.
How do you transcribe audio to text in French?
The common free methods all accept French: Word (French from France and from Canada), Google Docs voice typing, Voice Memos on the iPhone, Pixel’s Recorder, Live Transcribe on Android, Whisper and most online services. Choose the language before starting the transcription, with the --language French option for Whisper, because some services cannot reprocess a recording in another language afterwards. In Google Docs, punctuation voice commands only work in English, so full stops and commas are added by hand.
How do you transcribe audio to text on an iPhone?
With the Voice Memos app, on an iPhone 12 or later: it transcribes what it records in a set list of languages that includes English and French, and you can watch the text live by swiping up from the top of the waveform. After recording, open the recording, tap the options button, then Copy Transcript to take all of it, or View Transcript to select only part. The feature is not available in every country or region, and Apple mentions no separation of speakers.
Sources
- Microsoft Support, Transcribe your recordings (Word for the web), checked 1 November 2026support.microsoft.com
- Google Docs Editors Help, Type & edit with your voice, checked 1 November 2026support.google.com
- Apple Support, View a Voice Memos transcription on iPhone, checked 1 November 2026support.apple.com
- Google Pixel Help, Create, edit & manage transcriptions (Recorder), checked 1 November 2026support.google.com
- Google Pixel Help, Use speaker labels in a recording, checked 1 November 2026support.google.com
- Android Accessibility Help, Use Live Transcribe, checked 1 November 2026support.google.com
- OpenAI, Whisper GitHub repository (licence, installation, models), checked 1 November 2026github.com
- TurboScribe, Pricing page (free tier), checked 1 November 2026turboscribe.ai
- TurboScribe, home page (“Powered by Whisper”), checked 1 November 2026turboscribe.ai
- Otter.ai Help Center, Conversation, import, and app limits on the Basic free plan, checked 1 November 2026help.otter.ai
- Otter.ai Help Center, Transcribe conversations in English, Spanish, French, German, Japanese, or Chinese (Simplified)help.otter.ai
- Otter.ai Help Center, Speech & transcription accuracy FAQhelp.otter.ai
- OpenAI Help Center, Uploading files and audio to ChatGPT, checked 1 November 2026help.openai.com
- CNIL, GDPR, chapter 4, article 28 (processor, in French)cnil.fr
- CNIL, Transferring data outside the EU (in French)cnil.fr
- Légifrance, French Criminal Code, article 226-1 (in French)legifrance.gouv.fr
- Légifrance, French Labour Code, article L1222-4 (in French)legifrance.gouv.fr
- CNIL, Recording or listening to phone calls: must you inform your customers? (in French)cnil.fr

