How to Transcribe Audio to Text for Free

The short answer

If you have a recording and need its words as text right now, open a browser transcription tool that runs the model on your device. Allow the microphone or pick the audio file, let the model download on first use, and you get one continuous block of editable text back, with a word count and how long the clip and the run took, plus copy and download buttons. No account, no API key, no quota, no upload, and nothing to install.

If you are speaking live and want words to appear where your cursor is — an email, a doc, a chat message — that is a different job called dictation, and your operating system already does it for free with nothing to download.

If the transcript has to be defensible — something you will publish, quote in an article, submit to a court, or enter in a medical record — free machine transcription is the wrong tool, and the tool's own page says so.

• Saved recording into a text file: an on-device browser tool.
• Live speech into a text field: your OS or word processor dictation.
• Anything with legal, clinical, or published consequences: a human.

What “free” actually costs here

There is no invoice here, but there is still a price, and it is worth naming rather than pretending the word “free” means costless.

What you do not pay: an account, an API key, a monthly minute allowance, a “first 15 minutes free” trap, or an upload. No server is doing the work, so there is no server bill to recover and no reason for anyone to monetise your audio.

What you do pay. First, bandwidth: the model is not embedded in the page but fetched on first use, so the first run pulls weights before anything happens. Your browser keeps that copy and reuses it later, which makes it a one-time cost rather than a per-use one — still a cost on a metered connection.

Second, your own machine. Inference runs on your CPU inside the tab, so the job takes wall-clock time, a laptop battery drops noticeably during a long one, and you cannot background it and forget it the way you can an upload — close the tab and the work is gone.

Third, memory, which is the real ceiling rather than a stated limit: the page holds the whole clip while it works, so how long a recording you can attempt depends on how much memory the tab can get. A short interview is unremarkable; a two-hour recording on a cheap laptop is a coin flip.

What the model actually does to your audio

The model behind free browser transcription is Whisper, and the useful mental model is not “speech recogniser” but “translator between two representations”. Your waveform becomes a picture-like array of audio features — how much energy sits at each frequency in each short slice of time. A network reads that array the way you read a page of text, and a second network writes English words one at a time, conditioned on everything it has already written. It is the sequence-to-sequence shape that powers machine translation.

That architecture explains behaviours a phonetic spellchecker would not predict. Because text is generated word by word against a prior over how English usually goes, the output leans fluent and idiomatic: clear speech containing an unfamiliar word tends to come out as a familiar word that sounds close, because that is the more likely sentence. It also means punctuation and capitalisation are written by that same generator rather than by a rulebook, so they are habits, not guarantees.

30-second chunks, and the seams between them

The other structural fact is chunking. The model was trained on a fixed window of audio, and a recording of any length has to be cut into pieces. This page processes your audio in 30-second chunks and concatenates the results, each chunk starting exactly where the previous one ended. An hour of speech becomes 120 separate model calls.

The consequence is the most common complaint about long transcripts, and it is a property of this tool rather than a bug you caused. Because the chunks do not overlap, a word that straddles a seam can end up in neither window or in both, which is why an otherwise excellent paragraph contains a duplicated phrase and a dropped syllable at the same places every time you run the same file. Where one word carries the meaning — a dosage, a name, a negation — split the file at a pause first so a seam lands somewhere harmless.

English audio, and why that is a choice

This page runs an English-only Whisper checkpoint. Feed it English speech and it is built for the job; feed it another language and the output will not be a transcript of that language.

That is not a limitation to apologise for. Whisper comes in two variants: one trained to transcribe English, one trained to handle many languages and optionally translate them. The multilingual checkpoint spends part of its capacity on that breadth, and for English input the dedicated one is the better pick — it does not have to decide which language it is hearing before it decides what was said, and it is not tempted to quietly translate.

One corollary for mixed-language audio: a sentence of another language inside an English recording is exactly what a specialised model handles badly. Expect English words that sound like the foreign phrase, and transcribe those passages by hand.

What this tool does not do

Assume nothing here, because all three of these are things readers believe they are getting.

No speaker labels. Whisper does not do speaker diarisation, and this tool does not bolt it on. If three people talked, you receive one paragraph of text with no indication of who said what and no “Speaker 2” markers to clean up. For a multi-voice recording the fix is manual: leave a clear pause between speakers and label the sections while you edit, or transcribe each voice from a separate track if your recording has them.

No timestamps. You will not get timecode lines, word-level timings, or a clickable timeline that jumps to the audio. So you cannot cut a video, cite a moment in a deposition, or line up subtitles from this output without finding each position yourself. If timecodes are part of the deliverable, this route gives you the words and you supply the times.

No promise about punctuation or capitalisation. You may well get commas and sentence starts, but what is guaranteed is one continuous block of plain text, which leaves paragraphs, question marks and capitalised names as your responsibility. For one paragraph of monologue that is a five-minute cleanup; for eight thousand words of dialogue it is a real editorial job, and the missing paragraph breaks are what people never budget for.

Getting a transcript you can actually use

Free transcription is a lottery whose only winning ticket is good audio. Everything below is about the input, because the model is fixed.

Before you record

Get the microphone close and leave it there. A phone twenty centimetres from the speaker's mouth beats the same phone at the far end of a table, because proximity is what buys a clean signal over everything else in the room.

Choose the room, not just the mic. A hard-walled, empty, tall-ceilinged room is the worst environment there is: every consonant arrives at the microphone two or three more times, late and smeared, and the model has to commit to one interpretation of a sound that is genuinely ambiguous. A carpeted room with curtains and furniture records better, and closets and cars are absurdly good.

Kill the noise floor before you need it: fans, air conditioning, refrigerator compressors, open windows onto traffic, a running extractor. You will not hear them while you are busy talking; the model will.

One speaker at a time, and say so out loud. Overlapping speech is not two streams to a microphone, it is one waveform, and a sequence generator can only write one line of text, so it favours whoever is louder and loses the other person completely.

Sourcing audio you did not record

Grab the best available copy. An original file from the recorder, camera or platform beats a screen recording, which beats a re-encode: every generation of compression throws away the fine consonant detail the model relies on.

Be careful with phone-quality audio. Telephone bandwidth is roughly 300 to 3400 hertz, which keeps vowels comfortably and discards much of what separates “f” from “s”, “p” from “b”, “k” from “t”. Transcripts of calls come out fluent and wrong, which is the worst failure mode, because they read fine.

Video with music or a room tone under dialogue is harder than it looks: the model was trained on speech, not on speech mixed under a soundtrack, and a loud bed of music tends to produce confident nonsense.

For jargon, keep a short glossary beside you. Unusual names, product code names, drug names, place names and acronyms are the tokens the model has seen least often, so it substitutes the frequent word nearest in sound. Expect a name pass and plan for it.

After the transcript appears

Treat it as a first draft typed at impossible speed. Paste it into your editor, then do three passes rather than a read-through: numbers and dates, proper nouns, and anything sitting near a 30-second mark.

Check digits and names first, in that order. They break before anything else and they break silently — a wrong figure in a summary you publish is your error, not the model's, and there is no confidence score attached to the word to warn you.

Fix the structure last: break the single block into paragraphs at real topic changes, add question marks where someone asked something, and remove the duplicated phrase you now know to look for.

Free routes on a Mac, on a PC, and online

The answers people search for here, kept separate from the browser tool above: on both desktop platforms you already have free speech-to-text installed, and for some jobs it is the better choice.

On a Mac

macOS ships system dictation, switched on in the Keyboard settings and callable anywhere text can be typed — Notes, Mail, a browser field, most editors. Click somewhere, press the configured shortcut, speak, press it again. The same feature exists on iPhone and iPad inside the keyboard.

That route is for live speech into a text field and it is not a transcription tool: it types as you talk, it does not read an existing audio file, and if you go quiet it may simply stop listening. For a recording that already exists, use the browser tool or your word processor.

On a PC

Windows has voice typing in any focused text field, opened with the Windows key and H. It dictates your words into that field and takes punctuation by voice command rather than hearing it out of your phrasing.

Same caveat as macOS, and the same reason: this is dictation, not transcription. Windows does not ship a free “give me this audio file as text” utility, which is the gap the on-device browser tool fills.

Online

The browser route runs on any machine with a current browser, including Chromebooks and Linux desktops — the practical answer where the operating system has no dictation worth using or the machine is locked down.

Keep two things straight about “online”. It means in a browser tab, not on a server — the audio is decoded and transcribed locally. And the first run needs a connection, because that is when the model arrives; later runs reuse what your browser kept.

On a work machine, note that the page asks for microphone permission and policy may block it; the file route never touches the microphone.

Dictation, word processors, and cloud services

Three routes compete with the on-device tool, and each has one honest tradeoff.

Operating-system dictation

Fastest for short bursts, free, already installed, and it puts words directly where you are working. Unsuitable for anything you already recorded or cannot re-perform — you cannot dictate a podcast you captured yesterday.

Word-processor dictation modes

Mainstream word processors and note apps expose their own dictate feature, which is the same live-speech idea with formatting commands attached. Useful for drafting by voice, equally useless on an existing file. Some of them run through the vendor's online services rather than your device, so if the content is sensitive, find out what the feature does with your microphone before using it.

Uploading to a third-party transcription service

This is the classic “free online transcription” offer: usually an account, usually a signup allowance, and in exchange for convenience your recording is copied onto somebody else's infrastructure, processed there, and retained there for however long their policy says — a window that typically survives your deleting the local file.

For a public interview recording that is often acceptable. For anything containing a patient's name, a client's confidential remarks, or an interviewee who agreed to be recorded and not republished, it is a real decision and not a detail. The on-device route exists precisely so the question never has to be asked: the audio goes nowhere, so there is no retention period to read about.

Cloud services do buy you things — speaker labels, timecodes, very long files, and processing that does not tie up your laptop. Just do not buy them for a file you should not be sending.

Step by step with the browser tool

What the workflow looks like end to end, with the input rules stated plainly so nothing surprises you.

1. Load the page and let the model arrive

Open the tool and start a job. On first use the page reports the model download and you wait while it fetches the weights; later visits reuse what your browser already has. Nothing else installs and there is nothing to sign up for.

2. Get audio in

Two inputs. Record with the microphone in the page, which captures locally and hands you text back — the route for a rehearsal, an interview you are sitting in, or a note you would rather speak than type. Microphone capture produces WebM.

Or pick an existing file. There is no fixed list of supported formats, because the page holds no whitelist: whatever your browser can decode as audio or video can be handed to the decoder, and anything it cannot decode simply fails. A file your browser will play is a reasonable prediction of a file this page can read; try one short clip before committing a long one.

Either way the audio is resampled inside the page to 16 kHz mono — one channel, 16,000 samples a second, which is 960,000 samples in a minute — before the model sees it. That is the resolution Whisper works at, so a 48 kHz stereo master buys the transcript nothing.

3. Run it and read the result

Start transcription and the page walks the audio in 30-second chunks. The result is one continuous block of plain, editable text, with a line under it counting the words and reporting how long the clip and the processing took. Two ways to get it out of the page: a copy button, and a download button that saves the transcript as a .txt file. There is no subtitle, timecoded or word-timed export.

Paste into your editor as soon as the text lands rather than leaving it in the tab: the transcript lives only in that page, and a closed tab means a lost transcript.

4. Know your ceiling before you commit

The binding constraint on length is memory rather than a stated cap, because the entire clip is held while it is processed, so what works depends on your machine. If a file matters, transcribe it in sections and get one piece at a time instead of gambling on the whole job.

When to pay a human instead

Free machine transcription is a drafting instrument. Three categories of output should not be drafted.

Legal and regulatory material: depositions, hearings, contracts, anything where the exact words may be read back to you. The reasons are structural rather than about quality — no timestamps to locate a moment, no attribution to establish who spoke, and no way to show what the recording said as opposed to what the model guessed.

Medical and clinical: dosages, patient histories, consent conversations. A substitution of one plausible word for a similar-sounding one is the normal failure mode, and this is the domain where that failure is measured in harm. Numbers and drug names are precisely the tokens machine transcription handles worst.

Published or quoted material: journalism, a book, anything in which a named person's words will appear. Quoting a machine's approximation of someone's sentence under their name is a factual error waiting to be found, and the correction becomes the story. The tool's own page says so: for production-grade accuracy on legal transcripts or comparable work, use a larger model or a professional transcription service.

One more case: a human second pass is often cheaper than your own editing time. If a free transcript costs you forty careful minutes per hour of audio, and your hour is worth more than that, you saved nothing.

Common problems and what they mean

Nothing happens and no model arrives: the first-run fetch failed, usually a network problem or an extension blocking it. Reload, disable the aggressive extension, retry.

The transcript is fluent garbage: the audio is quieter, noisier or more reverberant than the model can use, and it did what the architecture does — wrote plausible English. Re-record or re-source with better proximity; nothing on this page will rescue a bad room.

One phrase repeated twice and one word missing at regular intervals: 30-second chunk boundaries, exactly as described above. Split the file at a pause and rerun that stretch.

A long file fails halfway, or the tab dies: memory. The clip has to fit in the tab, so split it and run the pieces.

The file is refused: your browser cannot decode it. Re-export from the source app at a standard setting.

Punctuation you did not say: expected, because the generator writes its own habits. Fix it in the edit pass and do not read it as evidence of where sentences ended.

Gibberish where the audio is not English: this page's checkpoint is English-only, so use a page built for that language.

Silence from the microphone: browser permission denied, or the wrong input device chosen in system settings. Check both before assuming the page is broken.

Frequently asked questions