Comparison

Teneks vs MeetGeek

Most comparison pages are a feature table nobody can check. This one is two recordings, both tools, and the output of each — including the parts that do not flatter us.

How we tested, before what we found

Two real recordings: a 48-minute Estonian discovery call and a 27-minute English client call with a three-minute Estonian side-conversation in the middle. Each was processed twice, once by each tool. Participants and their commercial details are anonymised throughout.

The part you should push back on

There is no independent human transcript for either call, so we used our own output as the reference. Scoring a competitor against your own result is not a neutral benchmark, and we are not going to pretend otherwise. That is exactly why the three findings below are the ones we lead with: none of them depends on which transcript you treat as correct.

Three findings that hold with no reference transcript at all

Six speaker labels for three people

On the English call, three people were in the room. MeetGeek's export contains six distinct speaker identities, including one carrying speech that nobody said, and leaves roughly one turn in nine attributed to "Unknown speaker". This is a countable property of their own file — you do not need our transcript to check it. Teneks returned three speakers, by name.

On the Estonian call the position partly reverses: MeetGeek emits exactly three labels, and we emit three plus six short unattributed backchannels. We are not going to quote only the half that suits us.

A 106-word repetition loop

About one word in twenty-six of MeetGeek's Estonian transcript sits inside a degenerate repetition run — a 106-token wall of one word, and an 84-token wall of another. It lands in the closing minutes, the part a manager actually reads. Our worst run on the same audio is five repetitions of a word the speaker genuinely said five times.

This is the classic small-language speech-recognition collapse, and it is a defect against any reference, including none.

Three minutes of Estonian returned as invented English

While the client waited for a screen share, two participants spoke Estonian to each other for about three minutes. MeetGeek did not detect the language change. It produced fluent English sentences that were never spoken and attached them to a speaker who was not in the room — one Estonian word became "What's up? What's up? What's up?"

You do not need a reference transcript to identify a fabrication. This is the finding we would want a security or risk function to see.

Why this happens, and why a model update will not fix it

All three follow from the same architecture: an English-first pipeline with one declared meeting language and no identity resolution. Estonian is out of distribution, and a language switch inside a call sits outside the model of the problem entirely. That is why the gap against a strong English-first tool is incremental on an English call and categorical on an Estonian one — and why it is not the kind of gap that closes in someone's next release.

When MeetGeek is the reasonable choice

If your meetings run in English, in one language, with clear audio and a stable set of participants, and what you want is a clean readable summary rather than an evidence record, the difference between us narrows a long way. MeetGeek strips disfluencies deliberately and reads more smoothly than we do — our transcripts are verbatim, which is better for evidence and coaching and worse for skimming. On the English call its repetition behaviour was actually slightly better than ours.

The case for us is specific: multilingual rooms, hard audio, and a record something downstream is going to act on.

What we are not claiming

Not a defect-free transcript. We track our own artefacts — duplication where processing chunks meet, the occasional attribution slip, short backchannels we leave unattributed — and we fix them in public release notes. On the single most prominent company name in the English call, both tools were equally wrong. If you run this comparison yourself you will find those things, and we would rather you heard it here first.

Run it on your own call

The useful test is not ours. Take a recording that would break a generic tool — a hybrid meeting, a bad line, two languages in the room — and put it through both. We will help you set that up.

The seven reasons teams choose Teneks