SiliconSense
Unlock all · $26.87

YuE2-3B in Russian: an open model that sings Cyrillic

Open song-generation models usually stumble on Russian: you have to rewrite the lyrics in Latin script, and the vocals still carry a foreign accent. We rented a GPU, installed YuE2-3B and fed it Russian lyrics in ordinary Cyrillic — no transliteration, no language hints. Below are three tracks, the measurements and a blind intelligibility check.

Tested on 16 September 2026 on a rented NVIDIA A40. Rental cost about 40 cents. The lyrics were written by us specifically for this test.

In short

Listen for yourself

“Ne gasni” — dark alt-rock, male baritone

2:47 of audio generated in 119 seconds. Loudness −13.6 LUFS.

“Steklo” — synth-pop, female vocal

2:37 of audio in 118 seconds. Loudness −11.9 LUFS.

“Vzlyotka” — folk-rock ballad, male tenor

1:44 of audio in 89 seconds. Loudness −14.5 LUFS.

How we checked intelligibility

Judging by ear lets you forgive anything, so we handed the finished tracks to a speech recognizer blind — without showing it the source lyrics. It identified the language as Russian with 100% confidence and returned the words near-verbatim.

Written vs heard

“Kak posledniy tyoplyy bilet” (ticket) → heard as vzglyad (glance)
Vzdokh sentyabrya” (a sigh of September) → heard as Sdokh (dropped dead)
third track — recognised in full, no discrepancies

These are mishearings, not mis-singings: similar-sounding words confused, rather than noise instead of speech.

The numbers

MetricTrack 1Track 2Track 3
Duration2:472:371:44
Generation time119 s118 s89 s
Faster than real time1,4×1,3×1,2×
L/R correlation0,860,910,61
Loudness−13,6 LUFS−11,9 LUFS−14,5 LUFS
Truncationnonono

Channel correlation below 1.0 means the stereo is genuine, not the same track in both ears. VRAM under load was about 9 GB; the weights take 7.3 GB plus a 0.5 GB audio decoder.

How it compares with HeartMuLa

We tested another open model the same day — HeartMuLa 3B. The comparison is useful because it shows what “the model knows Russian” actually means.

HeartMuLa 3BYuE2-3B
Russian in Cyrillicdoes not workworks
Transliteration neededyesno
Accentforeignclean
Channelsmono48 kHz stereo
VRAM~20 GB~9 GB
LicenceApache 2.0CC BY-NC + permission

So it is a trade-off: HeartMuLa has the free licence but no Russian; YuE2 has the Russian but a stricter licence. Details on the former are in a separate write-up.

The licence: read it before you install

This is a case where the licence matters more than the measurements. YuE2 weights are released under CC BY-NC 4.0 — non-commercial — plus a separate permission from the authors.

Verdict

For Russian this is the first open model that needs no humiliating workarounds: no rewriting lyrics in Latin script, no living with an accent. The audio is genuine stereo rather than an imitation, and it generates faster than it plays. The barrier to entry is low: nine gigabytes of VRAM is enough.

An honest caveat: three tracks are a check, not a benchmark. We did not measure how the model handles long forms, unusual time signatures or rare genres, and we did not blind-test it against commercial services. We checked exactly one question — does it sing Russian intelligibly. It does.

Style is easier to get from a prompt
Running your own model means a GPU, a setup and a licence. If you need a specific sound right now, a ready style prompt is faster: our catalog has them by artist and genre.
Open the catalog

FAQ

Does it really sing Russian without transliteration?
Yes. We wrote the lyrics in ordinary Cyrillic, with no Latin script and no language hint in the settings. Pronunciation is clean, word stress is correct, there is no foreign accent. That is the key difference from HeartMuLa, where Cyrillic produced gibberish and the text had to be rewritten in Latin letters.
How did you verify the words are intelligible?
Not by ear. We ran the finished tracks through Whisper large-v3 blind — without giving it the source lyrics. The language was detected as Russian with 100% confidence and the words came back near-verbatim. The differences were ordinary speech-recognition errors on sung vocals.
What GPU do I need?
Under load the model used about 9 GB of VRAM, so a 12–16 GB card is enough. We rented a 48 GB A40 at $0.49/hour, but that much memory was never needed.
Is it fast?
Faster than real time by 1.2–1.4×: a 2:47 track was generated in 119 seconds. All three of our runs finished within a minute or two.
Can it be used commercially?
Not by default. The weights are released under CC BY-NC 4.0 — non-commercial — plus a separate permission from the authors. Personal use is free, and the authors state they take no fees or royalties on what you earn from your own tracks. Use of the weights by a company, however, must be agreed separately: there is no public price list.
Will it replace Suno?
For Russian this is the first open model where the question is even worth asking. But we will not rank it on quality: three tracks are a check, not a benchmark. And keep the licence in mind — with Suno you pay a subscription, here you negotiate with the authors if it is for business.

Related guides