Your audio never leaves your device
Private transcription that keeps your audio on your device
Confidential transcription that never uploads your recording. It runs in your own browser, so your audio never reaches our servers or anyone else's. Free transcription without a subscription, and no signup.
Takes a little extra time, and your audio still never leaves your device. Labels go by how voices sound, so treat them as a best effort: one person can be split across two or three labels, similar voices can be merged, and someone joining partway through may be folded into a speaker already talking. Giving a number usually helps. Check them before relying on them.
Specifying this number improves accuracy. Count people who really take part, whenever they join. Someone who only says a word or two is better left out.
Timestamps each word instead of each phrase, so short interjections like "Yeah" go to the right person. On a busy four-person recording this corrected about 1 word in 5; it changes little when people talk in long turns. Adds roughly 40% to the time on a computer, and almost nothing on a phone.
Phone audio loses the high frequencies that separate consonants, so they are put back before transcribing. Only phone-quality recordings are treated, and you are told when it happens. Also even out volume is for one case: someone far from the microphone and much quieter throughout. It cost accuracy on other recordings, so use it only if a quiet person came out badly.
Worth filling in. The speech model knows ordinary English, not your subject, so a word it has never heard becomes whatever sounds closest, and it usually gets that word wrong every single time it comes up. A sentence is enough, and the names and jargon in it are picked out and watched for. It cannot guess a word you do not mention, so name the people and the terms that matter.
Free, no signup. The first time you use it, it takes a minute or two to set up. After that setup is quick. Works with MP3, WAV, M4A, OGG, FLAC and AMR, and long recordings are handled in parts.
What "private" usually means, and what it means here
Almost every transcription service describes itself as secure. Read closely and the word is doing a lot of work. Secure normally means the upload is encrypted in transit, the file is encrypted at rest, and access is controlled. All of that is genuinely useful, and none of it changes the underlying fact: a copy of your recording now exists on infrastructure you do not control, governed by a policy that can be revised.
Private, as used on this page, means something narrower and stronger. The recording is never transmitted. There is no second copy, no retention window to reason about, no vendor breach that could expose it, and no policy change that could later repurpose it. Not because we promise to delete it, but because we never receive it.
The questions worth asking any transcription tool
Whether or not you use this one, these five questions separate meaningful privacy from marketing language. They are the questions we would want asked of us:
- Where does the processing happen?If the answer is anything other than "on your device", your audio is being transmitted. Everything else follows from this one.
- How long is the file retained after processing?Look for a specific duration. "Deleted promptly" and "as long as necessary" are not durations.
- Is my audio used to train models? Check whether this is opt-out rather than off by default, and whether the opt-out covers subcontractors too.
- Who are the subprocessors? Most services route audio through at least one other company. Each one is a separate place your recording exists.
- Can I verify any of this myself? Usually not, which is the real point. A server-side claim can only be audited, never observed. An on-device claim can be checked in your own browser in under a minute.
When uploading is genuinely fine
Being useful means being honest about where this tool is not the answer. Upload a conference talk, a podcast, a public webinar, or a lecture you plan to publish anyway. The material is not confidential, the cloud service will be faster, it will handle three hours without complaint, and its accuracy on messy multi-speaker audio will likely be better than a quantized model running in a browser tab.
On-device transcription earns its constraints in a specific situation: when the recording is something you could not defend having uploaded. A privileged client call. A therapy session. An HR complaint naming people who have not been told. A source who agreed to talk on the condition that the tape stays with you. In those cases the choice is usually not between this tool and a better one. It is between this tool and typing it out by hand.
Every request this page makes, by name
Telling you to watch the network tab is only useful if you know what you should see. Here is the whole list, so anything not on it is worth asking about.
- meetingscribe.net for the page itself: HTML, styles, fonts and scripts, the same as any website.
- cdn.jsdelivr.netfor the transcription engine, a pinned version of Hugging Face's transformers.js, and for the code that runs it in the browser. If you load an AMR file, a small AMR decoder comes from here too, because no browser has one.
- huggingface.co for the model files themselves, fetched once and then kept in your browser cache: the Whisper model you chose, and in Accurate mode the speaker-segmentation and voice-matching models as well.
- cloud.umami.is for cookieless analytics, which records that a page was viewed and that events like a transcription finishing happened. It records no file names, no audio and no transcript text.
- pagead2.googlesyndication.com for advertising, and only if you accept cookies. Decline and the script is never loaded at all.
Every one of those is a download to you. What is never in that list is an upload of your audio, your transcript, your subtitle files or your speaker names, because there is no server here that could receive them. That is also why the second check is the more convincing one: transcribe something short so the models cache, disconnect from the internet entirely, and transcribe again. It still works.
Getting good results on sensitive audio
Confidential recordings tend to be the hardest kind: a phone on a table, a speakerphone, a small room with hard walls. A few things reliably help.
- Choose the small model for difficult audio. The extra download is worth it when the recording is quiet, accented, or overlapping. Save tiny for clean single-voice dictation.
- Set the language explicitly rather than leaving auto-detect on, if you know it. Detection occasionally guesses wrong on the first few seconds and colors the whole result.
- Split anything long. Name the parts in order and select them together, and they are joined into one transcript. This is more reliable than pushing a single ninety minute file through one tab.
- Fix names in the editor, not the recording. Speech models mishear proper nouns most of all. Correcting them once in the page before download is faster than any re-run.
If you want the mechanics rather than the argument, how it works walks through the whole pipeline, names every network request the page makes, and shows you the two checks that prove the claim without trusting us.
Last reviewed: August 2026
Frequently asked questions
- Is my audio really not uploaded?
- Correct. The speech model downloads into your browser and the audio is processed on your device. You can confirm it yourself: open your browser developer tools and watch the Network tab, or transcribe one file, disconnect from the internet, and transcribe another. It still works offline, which nothing server-based could do.
- Do I need an account?
- No. There is no signup and no login. You open the page, drop in a file, and transcribe. There is deliberately no account system, because an account exists to link activity to a person and we do not want that link to exist.
- What does it cost?
- It is free, with no file limit and no account. The site is supported by ads, which load only if you accept cookies. There is no paid tier, and no feature is held back behind one.
- Are you HIPAA or GDPR certified?
- We do not claim any regulatory certification. The tool is designed for confidential workflows because your audio stays on your device, but you remain responsible for your own compliance obligations. If you need a signed data processing agreement, a commercial vendor is the right choice.
- What is the catch?
- The honest catch is capability. Long recordings are transcribed in parts automatically, but the whole file is still decoded into memory first, so about an hour per file is the comfortable ceiling. Add a one-time model download and a quantized model that a large server-side system will beat on difficult audio. You are trading some accuracy and convenience for the recording never leaving your machine.