Record a meeting (or drop in a recording you already have) and get a timecoded transcript plus draft minutes - summary, decisions, action items, risks, open questions and figures - built right here in your browser. The speech model runs in this tab, so the audio and the notes never leave your device, and you can export the minutes as Markdown or the transcript as TXT, SRT or VTT.
The first run downloads a 78 MB speech model and keeps it in this browser; after that a meeting starts transcribing immediately and works with the network off. On Chrome or Edge you can also capture the audio of a shared tab, so the other participants in a call are transcribed alongside your microphone. A 0.1 MB voice detector skips the silent stretches first, so the wait tracks the speech rather than the meeting length; the timecodes you get back still point at the original recording.
Need live dictation into a text box instead? Use the speech to text tool. Want the opposite direction? The text to speech tool reads text aloud.
What the minutes look like
A worked example so you can judge the output before downloading anything: a three-minute standup transcribed into six timecoded segments, and the notes the page built from them.
Transcript that went in (six spans of speech across 3:02, with the quiet stretches between them skipped):
Meeting notes
2026-08-29 · 3:02 · language detected automatically
Summary
- [0:07] We agreed to move the invoice export to the new service on Friday.
- [2:12] How do we handle invoices that were already exported by the old job?
- [2:48] Last month the old job failed on about 3% of invoices, so the new one has to be measurably better than that.
Decisions
- [0:07] We agreed to move the invoice export to the new service on Friday.
- [1:36] Mai needs to review the reconciliation report before we sign off on the cutover.
Action items
- [0:31] I'll prepare the rollback script by Thursday so we have a way back if the export misbehaves. (by Thursday)
- [1:36] Mai needs to review the reconciliation report before we sign off on the cutover.
- [2:48] Last month the old job failed on about 3% of invoices, so the new one has to be measurably better than that.
Risks and blockers
- [0:58] The blocker is still the sandbox account
- [1:03] We are waiting on finance to approve the 2,500 USD test budget.
Open questions
- [2:12] How do we handle invoices that were already exported by the old job?
Figures mentioned
- [1:03] We are waiting on finance to approve the 2,500 USD test budget.
- [2:48] Last month the old job failed on about 3% of invoices, so the new one has to be measurably better than that.
Every bullet is a sentence somebody said: the sections are filled by matching cue phrases such as "we agreed", "I'll", "blocker" and "by Thursday", which is why the deadline appears in brackets. A line carrying more than one cue can appear twice - the 2:48 figure is both a number and a commitment, so it lands under Figures and under Action items. The point is a draft you edit, not a finished document.
Meeting Notes Taker
Press record, hold your meeting, press stop: the recording is transcribed with a speech model that runs inside this browser tab, and the transcript is turned into draft minutes you can edit, copy or download. It suits a standup, a client call or an interview you want written up without handing the audio to a note-taking service.
What the minutes contain
The notes are assembled from the transcript in six sections. Summary picks the sentences carrying the most repeated content words. Decisions collects lines with decision language ("we agreed", "approved"). Action items collects commitment language ("I will send", "needs to") and, when the sentence names one, the deadline it mentions. Risks and blockers, open questions and figures mentioned follow the same idea for problems, questions and any money, percentage or count. Every line keeps its timecode, so you can jump back to the moment in the recording and check it. Because this is a text-analysis pass, not a model writing prose, it can only surface sentences that were actually spoken and cannot invent a decision nobody made. On a computer with WebGPU an optional button asks a small language model, also running in this tab, to rewrite the same notes as prose next to the extracted version rather than replacing it.
Three ways to get the audio in
Choose one at the top of the tool. Microphone only works in every browser that can record. Microphone plus a tab or window is the option for online calls: Chrome and Edge on a computer let you share a tab and tick "share tab audio", and the tool mixes that with your microphone so both sides are transcribed. Upload a recording takes an audio or video file you already have - a Zoom recording, a voice memo, an MP3 - and transcribes it the same way. Recording other people is your responsibility; ask for consent before you start.
The first run, and why the wait is shorter than the meeting
The speech model is OpenAI Whisper base, about 78 MB in its quantised ONNX form. The tool shows the real download percentage, then the browser keeps the files, so later meetings start with no wait and transcription still works offline. With WebGPU a recording transcribes several times faster than real time; on the CPU it runs at roughly the length of the audio it is handed. Long meetings are processed in 30 second chunks, so memory stays flat.
The audio it is handed is shorter than the meeting because most of a meeting is pauses. A 0.1 MB WebAssembly voice detector scores every ten milliseconds of sound, and only the spans where somebody speaks go to Whisper - three tenths of a second are kept either side of each span so words are not clipped, gaps shorter than six tenths of a second are never cut, and the pass is abandoned when it would save less than a twentieth of the recording. Each kept span remembers where it came from, so every timecode is converted back to the original timeline before you see it: on a test clip, four short contributions spread over 28 seconds came to 7.8 seconds of speech, and the timecodes landed with no measurable error. The pass makes transcription quicker, never more accurate, so it is a tick box - untick it and every second goes to the model.
Privacy, licences and credit
No audio, transcript or note is uploaded: the model files come down to your browser and the processing happens locally. The speech model is Apache-2.0 licensed and served from our own asset server at a pinned version, checksummed when it is published. The voice detector is rnnoise-wasm (Apache-2.0), a build of Xiph RNNoise (BSD-3-Clause), loaded at an exact version and checked against a known digest every time before it runs. The idea comes from meetily, an MIT-licensed desktop meeting assistant; that project is a Rust and Tauri application with no browser build, so this page is an independent implementation of the same job for the browser rather than a port of its code.
Frequently Asked Questions
What does the Meeting Notes Taker do?
It records a meeting (or reads a recording you upload), transcribes it with a speech model running inside this browser tab, and builds draft minutes from the transcript: summary, decisions, action items with any deadline they mention, risks, open questions and figures. Every line keeps its timecode, and nothing is uploaded.
Is my audio uploaded anywhere?
No. The model files are downloaded to your browser and the recording is processed locally, which is why the tool keeps working with the network disconnected. There is no account and no server step.
Can it capture the other people in an online call?
On a computer running Chrome or Edge, yes: choose the microphone-plus-tab option, pick the meeting tab when the browser asks, and tick "share tab audio". That audio is mixed with your microphone. If the share has no audio track the tool says so and records your microphone only.
How long does transcription take?
With WebGPU it is several times faster than real time. On the CPU it runs at roughly the length of the audio the model is given, which is not the length of your meeting because the silent stretches are left out first - a 30 minute call with ten minutes of speech is about a ten minute wait on the CPU. The status line names which mode your browser is using, and long recordings are processed in 30 second chunks so memory stays flat.
What is the "skip the silent parts" tick box doing?
Before transcribing, a 0.1 MB voice detector keeps only the spans where somebody is speaking, plus three tenths of a second either side so words are not clipped. Gaps shorter than six tenths of a second are never cut, and the pass is dropped if it would save less than a twentieth of the recording. The timecodes you see are converted back to the original recording, so they still match the real meeting. It saves time, not accuracy - untick it and every second goes to the model.
Could the silence pass drop something that was said?
It can, and that is why it is a tick box rather than a fixed behaviour. A voice detector judges what sounds like speech, so a whisper or someone talking far from the microphone is the case where it could be wrong. The padding and the refusal to cut short gaps are there to make that unlikely; if a recording is quiet or breathy throughout, untick the box and transcribe the whole thing.
Why does the first run download 78 MB?
That is the speech model itself - the quantised ONNX build of OpenAI Whisper base. It is fetched once from our asset server, then kept in the browser cache, so later meetings start immediately and work offline.
Which languages can it transcribe?
Whisper is multilingual; the language list covers English, Vietnamese, Chinese, Japanese, Korean, Spanish, Portuguese, French, German, Italian, Dutch, Russian, Hindi, Indonesian, Thai, Arabic, Turkish and Polish, plus automatic detection. There is also a checkbox to write the transcript in English instead of the spoken language.
Does it label who said what?
No. There is no speaker separation, so the transcript is one stream of timecoded lines. If you need names in the minutes, add them while editing the exported Markdown.
Can the minutes invent something that was not said?
The default minutes cannot: they are selected sentences from your own transcript, matched by cue phrases, so every bullet is a line somebody actually spoke. The optional language-model rewrite is different - a small model paraphrases those notes, and it appears as an extra block next to the extracted version so you can compare before sending it on.
What can I export?
Minutes as Markdown (.md), the transcript as plain text with timecodes, and subtitle files as .srt or .vtt for dropping onto a video timeline. There is also a copy button for the minutes.
How is this different from the speech to text tool?
The speech to text tool is live dictation into a text box using the recognition built into your browser. This page captures a whole meeting, processes it afterwards entirely on your device, and produces a document with sections and timecodes.