Takeaway. The model is not the variable. Almost every notetaker in this category runs the same class of speech model, and the thing that decides your transcript quality is how far the speaker's mouth sits from the microphone.
The room roughly doubles the error rate. On the AMI meeting corpus, Whisper Large V2 scored 16.9% word error rate on headset audio and 36.4% on a single microphone placed in the room. Same meetings, same model.
Speaker separation is the second failure, and it is worse than the first. A transcript that gets the words right but attributes them to the wrong person produces action items assigned to whoever spoke loudest.
Bot-based tools have nothing to join. Any notetaker whose capture method is "a participant joins your call" has no in-room mode at all, whatever the marketing page says.
Test with your worst room, not your best one. A quiet two-person office proves nothing about a six-person room with a whiteboard, a glass wall and an air conditioner.
Most AI note takers were built for the video call. A bot joins as a participant, takes the clean per-speaker audio stream the conferencing platform already separates for it, and writes a summary. Move that same tool into a conference room and two of its three advantages disappear: there is no call to join, and there is no per-speaker stream. What is left is one microphone somewhere on a table, picking up four voices, a projector fan and whatever is happening on the other side of the glass.
That distinction decides which tools are even candidates, and it explains why a notetaker that people praise for Zoom calls can produce an unusable transcript from a client visit. The rest of this guide covers what the room does to the audio, how to set up capture so the model has a chance, what to test before you commit, and what consent law expects of you once a recording device is on the table.
Why a room is harder than a video call for an AI note taker
The clearest published measurement of this gap comes from the model most local notetakers run. The AMI Meeting Corpus is 100 hours of real meetings recorded twice at once, through headsets worn by each participant and through microphones sitting in the room. OpenAI benchmarked Whisper against both. In Table 2 of the Whisper paper, Large V2 scores 16.9% word error rate on the headset audio and 36.4% on a single distant microphone. Identical meetings, identical speech, one variable.
Three things in a physical room cause that.
Distance. Sound obeys the inverse square law: each doubling of the distance between mouth and microphone drops the direct signal by about 6 dB, while the reflected sound bouncing off the table, the window and the ceiling stays roughly constant. A person sitting at the far end of a four-metre table is not quieter in a way you can fix with gain. They arrive with a worse ratio of speech to room.
Overlap. People in a room interrupt each other far more than people on a call do, because there is no half-second of network latency to punish them for it. Two voices hitting one microphone at once are a single waveform, and separating them again is a research problem rather than a settings toggle.
Speaker attribution. Microsoft built a benchmark specifically for this. The NOTSOFAR-1 challenge, which ran as Task 2 of CHiME-8, recorded 315 short meetings across around 30 conference rooms so that researchers could work on diarization and recognition from one far-field device. It scores speaker-attributed and speaker-agnostic error separately, because getting the words right and getting the name right are two different failures, and in a room both happen at once.
| What the tool receives | Video call | The same meeting in a room |
|---|---|---|
Audio streams | One per participant, already separated | One mixed waveform |
What the tool receivesAudio streams Video callOne per participant, already separated The same meeting in a roomOne mixed waveform | ||
Speaker labels | Supplied by the platform | Inferred from voice alone |
What the tool receivesSpeaker labels Video callSupplied by the platform The same meeting in a roomInferred from voice alone | ||
Mouth-to-microphone distance | 30 to 60 cm | 1 to 4 m |
What the tool receivesMouth-to-microphone distance Video call30 to 60 cm The same meeting in a room1 to 4 m | ||
Background noise | Each person's own room, gated | Shared, constant, unmuteable |
What the tool receivesBackground noise Video callEach person's own room, gated The same meeting in a roomShared, constant, unmuteable | ||
Overlapping speech | Suppressed by latency | Frequent |
What the tool receivesOverlapping speech Video callSuppressed by latency The same meeting in a roomFrequent | ||
Tip: When a vendor quotes a transcription accuracy figure, ask which of those two columns it was measured in. Almost all of them come from the left one.
How to set up capture for an in-person meeting
Hardware and placement buy you more accuracy than switching tools does. Four steps, in the order that matters.
Step 1: Put the microphone in the middle, not in front of one person
A laptop on the table records the person behind it beautifully and everyone else badly. Move the capture device to the geometric centre of the seated group, off any surface that resonates, and away from the projector and the air vent. If the room is longer than about three metres (10 feet), one central device is not enough and you want either a second recorder at the far end or a dedicated conference microphone.
Step 2: Remove the three noises the model cannot separate
Continuous broadband noise is the one that hurts most, because it fills the gaps between words where the model looks for boundaries. Close the door, turn off the air conditioning for the hour if you can, and move the recorder off any table that carries tapping and cup noise. Paper shuffling directly next to the microphone will cost you more words than a conversation in the corridor.
Step 3: Open the meeting by naming who is in the room
Say the names out loud at the start, in seating order, with a short pause between them. This costs fifteen seconds and does two jobs at once. It gives you an anchor for correcting speaker labels afterwards, and on tools with speaker identification it gives the model a clean sample of each voice before anyone starts talking over anyone else. It also puts the recording on the record, which matters for the consent section below.
Step 4: Transcribe, then correct the first five minutes by hand
Recognition errors cluster at the start, while the model is settling on accents, names and vocabulary, and those early minutes usually carry the agenda and the participant names that the summary is built from. Fix them before you generate the summary rather than after. A meeting minutes template is a faster place to do that cleanup than the raw transcript pane, and it forces the decisions and owners into fields where a missing one is visible.
What to test before you trust a tool in the room
Run a real meeting, not a demo. Record your own team arguing about something for twenty minutes in the worst room you actually use, then check the following against the resulting transcript.
In-room evaluation checklist
The import test is the one people skip, and it is often the one that solves the problem. A phone lying on the table beside the speaker is closer to most mouths than a laptop at the head of it, and any tool that accepts an audio file lets you use the better recorder and still get the AI summary.
Where meetergo Log fits, and where it does not
meetergo Log is the notes app inside meetergo's suite, and it is a useful case for this article because its capture model is the one that translates to a room. It is a native desktop app for macOS 14.2 and later and Windows 10 and 11, with Linux listed as coming. Transcription runs on your own machine with Whisper or Parakeet across 40 and more languages with automatic detection, so no audio leaves the device and the app keeps working offline.
For an in-person meeting, the path that the product pages document is import rather than live room capture. Existing recordings drag straight into the app, with MP4, WAV and MP3 among eight supported formats, and the transcript is created locally. That covers the phone-on-the-table workflow described above, and it is the recommendation worth making, because the app's live capture is described as recording system audio from meeting tools rather than as a room-microphone mode. Push-to-talk dictation does use the microphone, which suggests live capture in a room is feasible, but meetergo does not advertise it as a supported in-person recording mode. Confirm it directly before you plan a client visit around it.
Speaker labels are the second condition. Diarization on Log sits on the Light plan rather than the free tier, and identifying recurring voices by name is on the roadmap rather than shipped. Since speaker attribution is the harder half of in-room transcription, treat the free tier as a transcript-and-summary tool for rooms and the paid tier as the one that answers who said what.
| Plan | Price |
|---|---|
Free | €0, local transcription, summaries, search, export, dictation |
PlanFree Price€0, local transcription, summaries, search, export, dictation | |
Light | €9.90/user/month, adds speaker diarization and custom templates |
PlanLight Price€9.90/user/month, adds speaker diarization and custom templates | |
Suite | €29.90/user/month, adds the rest of the app suite |
PlanSuite Price€29.90/user/month, adds the rest of the app suite | |
Premium | €49.90/user/month |
PlanPremium Price€49.90/user/month | |
There is also a cloud notetaker that joins meetergo connect, Google Meet, Teams and Zoom as a bot, hosted in the EU and included from the Light plan at 10 recaps a month. It is the wrong instrument for a physical room, for the obvious reason that there is no call for it to join, and it is listed here so the two capture modes do not get confused.
Transcribe meetings on your own machine, with no bot and no upload.
Transcribe meetings on your own machine, with no bot and no upload.
If you want the wider field rather than one tool, the comparison of AI notetakers by capture method and the longer AI note taker roundup both sort tools by whether a bot is involved, which is the first filter that matters here. For pricing across the suite, see meetergo's plans, and for where data is processed, the security overview.
Recording consent in a room full of people
A bot in a video call announces itself. A laptop transcribing silently in a meeting room does not, and that difference is legal, not cosmetic.
In the United States, federal law permits recording with the consent of one party to the conversation under 18 U.S.C. § 2511(2)(d), but several states require every party to consent. California is the strict example: Penal Code § 632 covers confidential communications and expects all parties to agree. If the people in the room are in different states, the safe reading is the strictest one that applies.
In the EU, a recording of identifiable people is personal data and needs a lawful basis plus transparency about what happens to it, which the EDPB's guide for small organisations sets out in plain terms. Where the transcription runs changes the answer materially: a model running on your own laptop involves no transfer at all, while a cloud notetaker adds a processor you must name. Our guide to GDPR-compliant AI meeting notes covers the documentation side in more depth.
Careful: "I'll just take notes" is not consent to an automated recording, and a disclosure buried in a calendar invitation is not a disclosure to the client who walked into your office. Say it in the room, before the first substantive sentence.
Mistakes that cost you the transcript
- Putting the laptop in front of the chair. It records the host clearly and the room badly, which is the exact inverse of what you need.
- Judging a tool on a two-person meeting. Diarization on two voices is close to solved. Six voices in a reverberant room is where the tools separate.
- Trusting the summary without reading the transcript. A summary built on misattributed speech looks fluent and assigns the wrong action items to the wrong people.
- Recording the whole day. Long unstructured audio produces long unstructured summaries. Record the meeting, stop the recording when it ends.
- Ignoring the export format. If the transcript cannot leave the tool as Markdown, PDF or JSON, your meeting record is hostage to a subscription. Meeting minutes software is worth comparing on this alone.
For structuring what you do with the output, the Cornell note-taking method works as well on a machine transcript as on a paper one, and the meeting agenda templates give the model an outline to summarise against.
FAQ
Can an AI notetaker work without any internet connection?
Yes, if transcription runs on your own hardware. Tools that process audio on the device rather than in a cloud keep working in a basement conference room or on a client site with no guest WiFi. Tools built around a bot joining a call cannot work offline at all, because the capture itself depends on the network.
What is the best microphone setup for transcribing an in-person meeting?
One device at the centre of the seated group, off a resonant surface, with the air conditioning off. For rooms longer than about three metres, add a second recorder at the far end and merge the files afterwards. Individual lapel or headset microphones beat any shared device, which is exactly what the benchmark gap between headset and room audio shows.
Will an AI notetaker label who said what in a room?
Sometimes, and less reliably than on a video call. On a call the platform hands the tool one audio stream per person. In a room the tool has to infer speakers from voice alone, which is why speaker attribution is scored as a separate error type in meeting-transcription research. Expect to correct labels, and check whether diarization is on the plan you are buying.
Can I record on my phone and transcribe it later?
That is often the better setup. A phone on the table sits closer to most speakers than a laptop at the head of it, and any notetaker with file import will transcribe the recording afterwards. Check the supported formats before the meeting rather than after. Our overview of audio transcription covers the file route.
Do I have to tell people in the room that I am recording?
Assume yes. US federal law allows one-party consent, several states require all parties to agree, and the EU treats a recording of identifiable people as personal data that needs a lawful basis and transparency. Saying it out loud at the start costs nothing and removes the question. See also our notes on AI transcription and free transcription software.
Turn a room recording into minutes without sending the audio anywhere.
Turn a room recording into minutes without sending the audio anywhere.
