Transcribe audio to text, free
Choose online processing without a speech-model download, or keep audio on this device. Review your transcript, then request a summary with decisions and action items.
Checking online availability… You can choose On this device in settings below.
Online sends audio to the service named before upload (Cloudflare Workers AI or Groq) for speech recognition. On this device runs a downloaded speech model in your browser without uploading audio. Choose a mode your recording permissions and organisation allow; an AI brief separately sends transcript text to an online AI service. See the privacy policy for processing details.
Online file transcription uploads the original recording to private Supabase storage so the service can send it to the selected speech-recognition service. The source is removed after processing, failure or cancellation when possible. Service access to uncollected audio and online results ends after 24 hours. Scheduled cleanup removes remaining copies; outages or failed deletion can delay physical removal. Download results you need to keep. Read the privacy policy before uploading private material.
Getting a transcript worth reading
Check the recording as well as the model
A transcript can contain errors even when the speech sounds clear. Check the selected language, listen to the recording and compare the text with what was actually said. Pay particular attention to names, numbers and missing or added words.
Before a new recording, make a short sample and check that each participant is audible. Try another microphone position if a voice is quiet or distant. If you have separate participant recordings, you can transcribe those individually and label them yourself.
Listen for overlapping speech, background noise and echo. For future recordings, try a quieter location and compare another short sample. These checks help you assess the source; they do not guarantee an accurate transcript.
On this device: which quality setting to pick
Faster uses Whisper base, a roughly 73 MB download. More accurate uses Whisper small, roughly 237 MB. Both are cached by your browser after downloading, subject to its storage and cache settings.
Start with a short section in the language spoken. If the result is not useful, try the other quality setting and compare both transcripts with the recording. The larger model needs a larger download and still requires review; this page does not establish an accuracy improvement for your particular audio.
Neither setting guarantees correct words. If a section remains unclear, review it manually or ask a participant to clarify it instead of treating a fluent transcript as proof of what was said.
Online processing and recovery
Online sends the original file for processing without first decoding the whole recording or downloading a speech model on your device. Upload speed, file length and available capacity affect the wait. Check the limits shown by the tool; free online capacity is shared. If the response is interrupted, check the saved job rather than uploading the same file again.
On this device: how long it takes
The recording is transcribed about five minutes at a time, each part in 30-second windows with overlapping audio. Review the joins as well as the rest of the text. On supported devices the model can use WebGPU; otherwise it uses WebAssembly. The page tells you which one it is using, how far it has got and roughly how long is left, measured on your machine as it goes.
Because it is your machine doing the work, the time scales with the length of the recording, the selected model and the resources available. If progress is slow, try closing other demanding apps or tabs. The page requests that the screen stay on where the browser allows it. Processing time varies; there is no fixed completion time.
Picking the language
Online and On this device can offer different languages. Choose from the list shown for your mode. The browser-model details below apply to On this device.
The same 21 languages as live meeting notes are listed. The tool does not detect the language for you. It starts on the language you used last time, here or on live meeting notes, or your browser's language, and it starts as soon as you choose a file, so check the language first. A Hindi recording left on English will be transcribed badly; if you notice, change the language and it starts again.
On a computer, Hindi, Tamil, Marathi and Arabic start on the more accurate setting, by default. Telugu, Marathi and Arabic are marked "low accuracy" in the picker. Review the output in the spoken language whichever setting you choose.
For a recording that switches languages, choose the language spoken most and review the sections in other languages carefully. The tool does not automatically change your selected language during a run.
What this does not do
It does not identify speakers or label their turns. If you need attribution, review and label the transcript yourself, or transcribe separate participant recordings when you have them. Spoken names do not create automatic speaker labels.
Nothing joins your call as a participant: it works on a recording you already have. Timestamps come with it, though — show the time of each line, or download SRT or VTT subtitles. For a meeting in progress there is a separate page — live meeting notes — which listens to the meeting tab and transcribes it during the call. Live meetings offer Online with Groq when available, with short batches and shared limits, or On this device recognition, which keeps recognition audio local and may need a language or model download first. Online transcription on this page is for an existing recording. For meeting-tab audio, use Chrome or Edge on a computer. Microphone capture is another option; available sources depend on your browser, device and permissions.
Getting the recording out of your meeting tool
In On this device mode, a supported MP4 recording can supply its audio track, so there is no need to convert anything first. MP3, M4A, WAV, AAC, OGG, OPUS and WEBM files work directly, and so do MOV videos, as far as your browser can read them. Online accepts the formats shown by its upload control, with its own size and duration limits.
Step-by-step for each platform: Google Meet, Zoom, and Microsoft Teams. If you are recording a call you are on, it is worth knowing where consent is legally required before you press record — in some places every participant has to agree, not just you.
Questions people ask
Does my recording leave this device?
It depends on the mode. Online uploads the original recording to private Supabase storage and sends audio to the provider shown before upload (Cloudflare Workers AI or Groq), without downloading a speech model to your device. On this device uses a downloaded browser model and does not upload audio. Review the disclosure before choosing Online; request an AI brief separately to send transcript text to an online AI service.
Is there a file size or length limit?
Online has per-file, daily shared-capacity and request limits, shown by the tool before upload. Availability can run out; an online allowance is not unlimited transcription. On this device has no service-imposed minute quota, but your device memory and browser decoding still limit what can run. Try a representative short section before a long recording.
Why is there a wait before the words start?
Online needs time to upload and process the recording; speed depends on your connection, file and available service capacity. On this device downloads its model and speech engine on first use. Cached files may be reused, but changing quality or clearing browser data can require another download. Neither mode promises a fixed completion time.
Does it work offline?
Online needs an internet connection throughout upload and while checking its result. On this device also needs a connection to load the page and engine; its first model/engine download is roughly 96 MB on Faster or 260 MB on More accurate. Cached model files do not make the entire site an offline application.
Can I get timestamps or subtitles?
Yes. When timings are available you can show the time of each line and download SRT or VTT subtitles. Estimated times still need review against the recording, especially around silence, overlapping voices and section joins.
Can it tell speakers apart?
Review and add speaker labels yourself when attribution matters. A spoken name in the transcript is not proof of who said the words. You can transcribe separate participant tracks when you have permission and align them by time.
Which file formats can I use?
The tool shows the formats accepted by Online before upload. On this device can read common audio files including MP3, M4A, WAV, AAC, OGG, OPUS and WEBM, and MP4 or MOV audio tracks when the browser can decode them. Format support is not a guarantee that a particular codec, long file or device will work.
Why is my transcript full of mistakes?
Check the selected language and listen for quiet voices, overlapping speech, noise and echo. For On this device, compare Faster and More accurate on a short excerpt if useful. Online can make errors too. Check names, numbers, negatives and missing or added words before relying on either result.
Does it detect the language automatically?
Choose the spoken language before selecting a file. The languages offered depend on the mode; do not assume a language supported locally is also offered online. For speech that switches languages, review those passages carefully rather than treating fluent text as proof of accuracy.
Where are online recordings and results kept?
The original upload is held in private Supabase storage for processing and sent to the provider shown before upload (Cloudflare Workers AI or Groq) for recognition. The service attempts to delete source audio after processing, failure or cancellation. Service access to uncollected audio and results ends after 24 hours; scheduled cleanup removes remaining copies, with delays possible during outages or failed deletion. After this browser confirms a successful local save, the service attempts to remove its result copy and retries failed deletion. Saved browser transcripts are separate: download anything you need to keep.
What if the connection drops or I cancel?
Use the saved online job to check its status instead of uploading the recording again. Stopping a wait is different from confirming that server processing stopped. A cancellation or unknown outcome may still use the reserved shared allowance. Keep the page open until your transcript has saved, or download a copy before leaving.
Can it transcribe a meeting that is happening right now?
Use live meeting notes for a conversation happening now. Live meetings offer Online with Groq when available, with short audio batches and shared limits, or On this device recognition with a first-use language or model download when needed. Select the audio source, tell participants and check permissions. Sharing a browser tab needs desktop Chrome or Edge; a phone microphone cannot capture a call in another app on that same phone. Online transcription on this recording page processes an existing file, not a live call.