Artificial intelligence

Five note-taking assistants: which understands Kazakh best

We dictated the same meeting in three languages and compared the transcripts from five assistants. Not all of them handle mixed speech.

For you: If your team switches between Kazakh and Russian, you will know which service to pick and what to check in every transcript.

Five note-taking assistants: which understands Kazakh best

Meeting transcription services promise to save an hour a day. No more writing minutes, hunting for who promised what, or replaying the recording. In Kazakhstan, though, these tools face a special test: mixed speech. In a typical meeting someone starts a sentence in Russian, drops in a Kazakh word and ends with an English term.

We tested five assistants on exactly that. We dictated the same meeting in three languages and put the transcripts side by side. Below is what we found, how to choose a service for your team, and what to check before trusting it with important conversations.

How we tested

Every assistant got the same meeting. The content stayed the same; only the language changed. That keeps the comparison fair: same voice, same pace, same topic. We looked at three things:

  • whether words were recognised correctly;
  • whether the meaning survived when the speaker switched language;
  • how names of people and places were spelled.

The last point matters more than it seems. If a name is wrong in the minutes, a task can end up with the wrong person.

The result: who coped

The gap showed up as soon as people switched between Russian and Kazakh mid-conversation. Mixed speech was handled best by the assistant that detects the language of each phrase first. It treats every phrase separately and only then recognises the words.

The other four behaved differently. They tried to “translate” Kazakh words into Russian and mangled the meaning. A Kazakh word comes out as a Russian word that sounds similar. The text looks clean, but it says something that nobody said. That is the worst kind of error, because you do not notice it while reading.

In single-language meetings the picture changes. All five did well, and the differences were small.

Why mixed speech is hard

A speech recognition model is usually tuned for one language. If you set it to Russian, it looks for a Russian word in every sound. When a Kazakh word comes along, the model does not know it and offers the closest Russian match. The neural network never leaves a gap; it always picks something.

Detecting the language of each phrase first solves part of the problem. It is not perfect either: if one or two Kazakh words slip into a Russian sentence, errors still appear. So you still need to check the output.

What does this mean day to day? A student records a lecture, a journalist transcribes an interview, a business owner writes up a meeting with a partner. In each case a misrecognised Kazakh word has to be fixed by hand, and the hour you saved is gone.

How to choose a service

Before you pay, look at these criteria:

  1. Language choice. Can you set the language manually, or does the service detect it phrase by phrase? For a bilingual team this is the main requirement.
  2. Proper names. How does it spell names and titles? Some tools let you add your own vocabulary.
  3. Where recordings are stored. Meetings often include business or personal details. Find out whose servers the audio goes to.
  4. Export. How easy is it to move the text into a document or a message?
  5. Free trial. The best test is a recording of your own meeting, not the vendor's demo.

A prompt for checking the transcript

Once the transcript is ready, give it to a chatbot and ask it to flag suspicious spots. Copy this prompt:

Below is a meeting transcript. People spoke a mix of Kazakh and Russian. List the words and phrases that may have been misrecognised or do not make sense. Separately, list every person's name and every proper name. Do not change the text, only flag it.
[paste the transcript here]

Keep the “do not change the text” line, or the chatbot will start “fixing” the transcript and add errors of its own. Check the flagged spots against the recording.

What this means for you and what to do now

The verdict is simple: for single-language meetings any of the five will do. If your team speaks two languages, pick a service with manual language choice and double-check proper names.

Steps you can take today:

  • Record one ordinary team meeting, the way you usually talk, switching languages.
  • Upload that recording to the free tiers of two or three services.
  • Compare the transcripts at the points where the language changes.
  • List your colleagues' names and company names and see which service got them right.
  • Before sending the minutes, check decisions and deadlines against the recording.

An assistant saves time, but the final word is yours. Checking the minutes of an important meeting once is far easier than fixing a task that went to the wrong person.

Ad Your ad hereA banner on the homepage, in stories and sections. Formats and prices on our Advertise page.

Comments

No comments yet. Be the first.