AI Meeting Transcript Cleaner & Summariser

A meeting transcript looks like a simple document, but it is really the output of a chain of systems: microphones, a speech recognition engine, a speaker separation step and a caption file format that was designed for video players rather than for people. Each link in that chain adds its own metadata, and most of that metadata is noise once the meeting is over and you just want to read what was said. This guide explains where that noise comes from, how caption files are structured, how timestamps and speaker labels can be removed safely, and how to turn a long transcript into notes that someone can act on, without sending a confidential conversation to a server you do not control.

How automated speech recognition turns a meeting into text

Automated speech recognition (ASR) starts by slicing the audio into very short frames, typically 10 to 25 milliseconds each, and converting each frame into acoustic features that describe how energy is spread across frequencies. Older systems passed those features through separate acoustic, pronunciation and language models. Most current engines, including the ones built into the big meeting platforms, use end-to-end neural networks that map audio directly to words or word pieces, then use a language model to choose the most likely sentence. The result is not a single answer but a best guess, and every word comes with a start time, an end time and often a confidence score.

That timing data is why transcripts are full of timestamps: the engine knows exactly when each phrase was spoken, and caption formats are built to show text in sync with a recording. Recognition is also why transcripts contain errors. Accents, people talking over each other, poor microphones, product names, acronyms and industry jargon all reduce accuracy, and a sentence can be grammatically perfect while still being wrong. A cleaned transcript is easier to read, but it is still only as accurate as the recognition underneath it, so names, numbers and commitments deserve a second look.

Speaker labels come from a separate process called diarisation, which answers the question “who spoke when?”. On Zoom and Microsoft Teams each participant usually has their own audio stream, so the platform can attach the display name of whoever was unmuted. When several people share one room microphone, the system has to cluster voices by sound alone and falls back to labels such as “Speaker 1” and “Speaker 2”. Both approaches can misattribute short interjections, especially when two people speak at once.

The anatomy of a caption file

Meeting platforms export transcripts in a handful of formats. Knowing which one you have makes it much easier to see what needs to be removed.

A WebVTT file begins with the word WEBVTT on the first line. It is followed by blocks called cues, separated by blank lines. Each cue can have an optional identifier, a timing line made of a start time, an arrow and an end time, optional positioning settings, and one or more lines of text. VTT also allows NOTE comment blocks, styling tags and voice tags that name the speaker. SubRip is older and simpler: every cue is numbered, the timing line uses a comma before the milliseconds instead of a full stop, and there is no header. YouTube’s SBV format is similar, with the start and end time separated by a comma on one line.

Because captions are designed to fit on a screen, each cue holds only a few seconds of speech. A one-hour meeting can easily produce more than a thousand cues, which is why a raw export is so tiring to read: every sentence is broken into fragments, interrupted by numbers and repeated names.

Timestamp payload parsing: why it is harder than it looks

Removing timestamps sounds like a single search-and-replace, but timestamps appear in many shapes. Hours may be missing (01:05), padded (00:01:05) or not padded (0:1:5). Milliseconds may use a full stop or a comma. Some exports wrap times in square brackets or parentheses, others put them at the end of a speaker line, and VTT can even embed word-level times inside the text, such as <00:00:03.200>.

The bigger risk is removing something that was actually said. In a sentence like “let’s meet at 3:30”, the time is part of the meeting content, not metadata. A careless pattern deletes it and quietly changes the meaning of the notes. This tool avoids that by only treating a time as metadata when it sits where metadata lives: in a cue timing line, alone on a line, inside brackets, at the very start of a line in a transcript where lines regularly start with a time, or straight after a speaker name. Times in the middle of a sentence are left exactly as spoken.

Timestamps are not useless, though. If you need to jump back to a moment in the recording, turn off “Remove Timestamps” and the tool will keep one clean start time at the beginning of each speaker turn instead of one on every caption line.

Speaker labels: from diarisation to readable turns

Caption formats repeat the speaker name on every cue, so a two-minute answer from one person can show the same name twenty times. The “Merge Repeated Speakers” option joins consecutive cues from the same person into one turn with a single label, which is how a human note-taker would write it. “Strip Speaker Tokens” removes the names altogether. That is useful when you want a plain block of text for a report, or when you are sharing notes outside the meeting and do not want to attach names to every opinion.

To recognise speakers, the tool looks for voice tags, bracketed names, Teams-style name and time lines, and short labels followed by a colon. A colon label only counts as a speaker if it appears at least twice or follows the “Speaker 1” pattern, so a one-off line such as “Note: the deck is attached” is not mistaken for a person. Diarisation errors in the original transcript cannot be fixed automatically, so if two people were talking over each other, check who said what before you rely on an attribution.

Verbatim, clean verbatim and edited transcripts

Professional transcribers distinguish between three styles. A verbatim transcript keeps every “um”, false start and repetition. A clean verbatim transcript removes those disfluencies but keeps the speaker’s words and order. An edited transcript goes further and rewrites sentences for grammar and flow. This tool produces clean verbatim: it removes common fillers (um, uh, er, erm, hmm, mm-hmm) and stutters such as “the the” or “w- we”, but it never rewrites what someone said. Words like “like”, “actually” and “you know” are left alone, because they often carry meaning and removing them automatically can change the tone of a sentence.

For legal proceedings, research interviews, HR investigations and anything where exact wording matters, keep the original export alongside the cleaned version. The original is the record; the clean version is for reading.

From transcript to summary: extractive and abstractive methods

There are two broad ways to summarise text. Abstractive summarisation, used by AI assistants, writes new sentences that paraphrase the meeting. It reads naturally but can invent details or merge statements from different people. Extractive summarisation selects the most important sentences that were actually spoken. It is less polished but every line can be traced back to the transcript, and it can run instantly in a browser without sending anything away.

The Summary tab in this tool is extractive. It scores sentences by how many of the meeting’s most frequent topic words they contain, then lists the strongest ones in the order they were said. It also uses language patterns to pull out likely decisions (“we decided”, “that’s agreed”, “let’s go with”), action items (“I’ll send”, “we need to”, “by Friday”) and open questions, and it calculates how much each speaker talked. Treat it as a strong first draft of your meeting notes, not as minutes approved by the participants.

Why local processing matters for meeting data

Meetings carry some of the most sensitive information an organisation has: unreleased plans, financial figures, customer details, performance discussions and personal opinions. Under data protection laws such as the GDPR and UK GDPR, a transcript that names people and records what they said is personal data. Many online transcript tools upload the file to a server for processing, which means another company now holds a copy of your meeting.

This tool works differently. The page loads once, and from then on every step, from reading the file to building the summary, happens in your browser’s memory on your own device. Nothing is uploaded, nothing is stored and closing the tab clears it. You should still follow your organisation’s rules on where transcripts may be kept, and remember that recording or transcribing a meeting usually requires telling participants, and in some places their consent.

A practical workflow checklist

1. Export the transcript from Zoom (cloud recording audio transcript, .vtt), Teams (Download as .vtt or .docx) or Google Meet (the transcript doc saved to Drive).
2. Open or paste it into the left-hand box. The format badge confirms what was detected.
3. Choose your options: keep timestamps if you need to find moments in the recording, strip speaker names if the notes are going outside the team.
4. Check names and numbers in the clean text, since recognition errors survive cleaning.
5. Read the Summary tab and edit the action items into your notes or task tracker, adding owners and dates where they are missing.
6. Copy or download the result, then delete raw exports you no longer need.

A transcript is only valuable if someone reads it. Removing the machine metadata turns an export that nobody will open into notes that people actually use, and doing it locally means the conversation stays where it belongs.