How we turned a 65-minute interview into a Ukrainian article

    Stengr team ·

    Pro Shcho? (proshcho.com) is a Ukrainian-language online magazine. Its section Pro Shcho Hovoryat publishes long interviews in full Ukrainian translation, so readers who don’t follow English can hear the whole conversation, not a summary of it.

    One of them is a conversation between Peter Thiel and Mathias Döpfner, CEO of Axel Springer, from his show MD MEETS. The video was released on September 24, 2026. It runs 65 minutes 29 seconds, in English, with two speakers. The Ukrainian version, Peter Thiel on the crisis of artificial intelligence, the decline of Europe and the struggle for the future, came out on October 4.

    Here is how it was made, with the numbers and the mistakes.

    What we used

    Not the Stengr Android app: the app records your voice, and this was a YouTube video. We used an internal tool for videos built on the same technology as Stengr. Speech recognition comes from Soniox, the engine that also runs live conversation translation in Stengr, and translation from an OpenAI model.

    The process

    1. The audio of the video was recognized with speaker separation, so every line was marked as Thiel or Döpfner. YouTube subtitles were not used.
    2. The text was split into segments at pauses. People don’t pause only at the end of a sentence: in this interview, 106 segment breaks fell in the middle of a sentence. When the text was assembled, the pieces were joined back into sentences and paragraphs.
    3. Each segment was translated into Ukrainian separately.
    4. The nine YouTube chapters became nine section headings, each with a link to its timestamp in the video.
    5. A person checked every segment against the English original, with AI help: each Ukrainian segment was compared with the English one, and the result was a list of proposed fixes. Nothing went in without the editor’s approval.
    6. The editor wrote the headline, the lead and the announcement. The rest is a full translation without cuts: about 8,080 words in nine sections.

    81 fixes, 17 of them changed the meaning

    The review made 81 fixes. Most of them were small, but 17 changed what the text said. The most instructive ones started in speech recognition, and the translation then carried the wrong word on with full confidence:

    • Recognition heard “tax” where the speaker said “tech stagnation”.
    • “Bend” instead of “ban the knife”.
    • “President” instead of “present”.
    • “Lefchin” for Max Levchin, PayPal co-founder.
    • “Linka” for Die Linke, the German left party.

    Translation added its own mistakes:

    • “The 21st century” where the speaker was talking about the 22nd.
    • A Russian word for “supersonic” instead of the Ukrainian one.
    • The host suddenly switched to the informal “you”. Ukrainian, like German or French, has a formal and an informal form, and the conversation used the formal one.

    What we learned

    The translation was accurate word for word, but it can’t see a recognition error. It translates a wrong word as confidently as a right one. That is why the most important part of the review was checking the text against the original audio and the English text, not polishing the style.

    Translation loses the speaker’s gender. Ukrainian verbs and adjectives change with the speaker’s gender. Thiel and Döpfner are both men, so it didn’t show here. In another interview, with a female host from The Economist, the translation gave her male forms in 11 places.

    Chapter titles need a separate look. They are translated apart from the conversation, without context, and sometimes shift the meaning.

    A full translation of someone else’s interview is a question of rights. Pro Shcho? names the original and links to the video, and the publisher can ask for the text to be removed.

    Without a human check, a text like this can’t be published: 17 mistakes that change the meaning in 65 minutes is too many for a reader who can’t check the original.

    What this means if you use Stengr

    The same rule works for your own recordings: speech to text gives a strong draft, and the words that matter deserve a second look.

    • Names are where recognition fails most often. In Stengr Projects, add your characters, places and rare terms to the names dictionary, and they are spelled the same way in every scene.
    • Stengr keeps the original next to the translation, so you can check a doubtful line against what was actually said.
    • For an important conversation, re-read the names, numbers and dates before you share the text.

    Try Stengr on Android

    Free plan to start, no sign-up, no ads.

    Get it on Google Play