Who is DimaTorzok?

Where you see this phrase

You may run into it when speech is transcribed automatically in:

• Telegram voice-message transcription;
• auto-captions in CapCut, Instagram Reels/Edits, DaVinci Resolve;
• interview, podcast, or lecture transcription via Whisper;
• .srt / .vtt subtitle generation for video.

What all cases share: there was a pause, silence, or music in the recording at that moment.

Why the AI writes it

Whisper was trained on roughly 680,000 hours of audio that already had subtitles — including YouTube and pirated releases. At the end of many of those videos, a line like “Subtitles by …” sits over silence. The model learned: silence at the end ≈ author credit. So when it hits empty audio, it confidently adds what it expects to see.

This is a model hallucination — fabricated text that wasn’t in the audio. The same bug appears in other languages: Turkish “Altyazı M.K.”, Czech “Titulky vytvořil JohnyX”, Arabic “ترجمة نانسي قنقر”, and more.

Debunking the “he hacked the AI” myth

Social media sometimes claims DimaTorzok “embedded a digital fingerprint in silence and scammed every generator.” That’s false.

There was no injection, hack, or fingerprint. Someone using the name DimaTorzok once made subtitles and signed their work — normal practice. Those credits accidentally entered Whisper’s training data along with millions of others. Less dramatic than the myth, but that’s how it works: training-data noise, not malice.

Was there a real DimaTorzok?

Probably yes — a pseudonym for someone who voluntarily subtitled others’ videos and signed them. Because that signature entered training data, the name now “lives” inside the model and surfaces for thousands of people worldwide. We describe a technical phenomenon and do not publish personal data or speculation about a specific individual (see disclaimer below).

How to remove it

Delete the credit line manually in your subtitle editor or transcript. If it keeps appearing:

1. Trim silence at the start and end before transcription.
2. Enable VAD (voice activity detection) to skip empty segments.
3. Post-process .srt / .vtt files and strip known hallucination signatures (“Subtitles by …”, «Субтитры сделал …», etc.).
4. For batch jobs, add filter rules for common Whisper tail artifacts.

This is a technical cleanup task — not “removing a virus,” but deleting a model artifact.