Our AI Wrote a Line Into Russian Songs That Nobody Sang. Here's What We Changed.
November 9, 2026 · 5 min read
One of these lines was never sung
The sung lines here are placeholders. The crossed-out line is real: it appeared, word for word, in our transcripts of Russian songs.
The short version
More than half of our Russian songs had a line nobody sang. 37 of 67 songs carried a subtitle credit like "Субтитры сделал DimaTorzok" in their lyrics — and in the videos made from them.
Whole verses were missing. Russian songs came back with about half as many words per minute as English or Spanish ones.
Now every Cyrillic song is heard three times. The best version of each 30 seconds is kept, invented lines are removed and a second model corrects the words.
1,022 words became 1,665. Measured on nine real Russian songs, old system against new. One song went from no lyrics at all to 234 words.
Try it with your Russian song
Upload it and you get the lyrics with timing and 7 reels in about a minute. If a word is still wrong, you can correct it yourself afterwards.
First song free, no account
How we found out
An artist who sings in Russian wrote to us: the text in his videos had mistakes. We looked at his songs and saw it at once. In one of them, half a minute of the transcript was a single word — a username — where a full verse should have been.
So we checked every Russian song analysed in the weeks before. It was not one unlucky upload. 55 % of them had at least one invented line, and a third only started showing lyrics 30 seconds into the song. English songs had no such lines at all.
Why a speech model invents subtitle credits
Speech models learn from enormous amounts of subtitled video. Russian subtitle files often end with a line that names who made them. The model has seen that line so many times that it has become its answer for "I hear sound, but I'm not sure there are words".
Music is exactly that situation. An intro, a drop, a bar where the beat is louder than the voice — and the model writes the credit instead of nothing.
In our case something made it much worse. We give the model a short hint in English so it keeps ad-libs like "yeah" and "ooh", and we add the song title. With a title in Cyrillic letters behind an English hint, the model tipped into subtitle mode for the whole song. One test song returned 4 real words with the hint and 232 without the title.
The three changes
Every song is heard three times. Once with our usual hint and the song title, once with no hint at all, and once with the hint but without the title. Each pass fails in different places. For every 30 seconds of the song we keep the pass with the most real words.
Invented lines are removed. "Субтитры сделал…", "Субтитры подогнал…", "Редактор субтитров…" and their relatives are recognised and taken out, also when they hang on the end of a real line.
A second model corrects the words. It writes the lyrics while the first model keeps the timing. This step was already running for English, Spanish, French and German. For Cyrillic songs it had been switched off by an old rule.
Before and after, on nine real songs
We ran the old and the new system on the same nine Russian songs, between 30 seconds and three minutes long.
| Before | After | |
|---|---|---|
| Words recognised | 1,022 | 1,665 |
| Songs with an invented line | 5 of 9 | 0 of 9 |
| Songs with no lyrics at all | 1 | 0 |
We also made sure nothing else moved: on the same audio, songs in English, Spanish, German and Japanese come out exactly as before.
What is still not perfect
A voice the model can't hear stays unheard.In three of the nine test songs the first 30 seconds are still missing, because the vocal sits under the beat there. Three passes help when one of them catches it. They don't help when none does.
That is what the new lyrics editor is for. Open your song, paste your text, and your words go onto the timing of the song — here is how it works.
FAQ
Songs in Russian, Ukrainian, Bulgarian, Serbian, Macedonian, Belarusian and Kazakh. They all share the Cyrillic script and go through the same three steps.
No. The change applies to songs uploaded from November 9, 2026. A song analysed before that keeps its lyrics until you correct them with Edit lyrics or upload it again.
No. A verse that is buried under the beat can still be missed, and names and slang can still be misheard. On our nine test songs the new system finds about 60 % more words than the old one, but not all of them. For the rest there is the lyrics editor.
A subtitle credit. Speech models learn from subtitled video, and Russian subtitle files often end with a line naming who made them. When the model is unsure what it hears — music without clear vocals, for example — it writes that line. It is a known behaviour of the Whisper models, not something in your song.
Lyrics that are actually in your song
Upload a song in Russian, Ukrainian or any other language. First song free, no account.
More guides
Tools for this
Find the Viral Radar for your genre
The best Songbrain songs in your genre, and where your own track would land.
