HeartMuLa 3B in Russian: testing the open Suno alternative
Open models that generate songs and covers came up in our community. Instead of guessing we rented a GPU, installed HeartMuLa 3B — a free model under the Apache 2.0 licence — and ran the same Russian lyrics through it nine times. Here is the result, with audio, numbers and cost.
In short
Cyrillic does not work. Four runs out of six collapsed after 5–38 seconds instead of three minutes. The one full-length take sings nonsense: the vocal transcript contains fragments of Arabic, English and Chinese, no Russian at all.
Latin script works. The same lyrics transliterated produced full-length tracks with recognisable words.
The accent is foreign, though. The model has never heard Russian: stress and softness drift, and by the end it wanders into other languages.
Listen for yourself
English lyrics — the model working as intended
Control run on their own example: steady structure, a full 2:35, tags “piano, happy”.
Russian lyrics in Latin script — the workaround that works
The same Russian lyrics written in Latin letters, tags “rock, male vocal, dark, cinematic, guitar”. A full 2:29.
What the recognizer heard in that track
“Gorod spit na dne… Fonari tonut v tishine… Ne gasni, ne gasni, ostansya so mnoy… Poka ne ostil let od svred… Dažde je sli drogi net”
The opening lines come through accurately, then pronunciation drifts — that is the accent in action.
What it costs
VRAM: about 20 GB, a 24 GB card is enough.
Speed: roughly real time — 2.5 minutes of music takes 164 seconds.
Setup: 21 GB of model files, deployable in an evening.
Our test: 1.5 hours of rental, nine tracks — about 75 cents.
Verdict
For Russian-language songs the model is not ready: transliteration with a foreign accent is not a reason to leave your current tool. But the licence is free, deployment is quick and English quality is decent. If the promised 7B version ships, or someone fine-tunes it on Russian, re-testing takes an hour and we will do it.
One useful find on the side: the same family includes an open lyrics recognizer that handles Russian vocals surprisingly well — we will look into it separately.
Meanwhile — prompts that work
While open models catch up, style is easier to get from a prompt. Our catalog has ready style prompts by artist and genre.
The model was trained on English, Chinese, Japanese, Korean and Spanish — Russian was not in the training data. In four of six runs generation collapsed after 5–38 seconds, and the one full-length take sang gibberish instead of the lyrics.
What does "transliteration works" mean?
Write the same Russian lyrics in Latin letters and the model sings the whole song, with words recognisable by ear. The pronunciation stays foreign though — stress and softness drift into a Slavic-Balkan accent.
What GPU do I need?
The model takes about 20 GB of VRAM, so a 24 GB card is enough. We rented a 48 GB A40 at $0.49/hour; the whole nine-track test cost around 75 cents.
Can it be used commercially?
HeartMuLa weights are Apache 2.0, so commercial use is allowed. The MuLaCover add-on for covers is non-commercial and needs written permission from the authors.
Will it replace Suno?
Not yet. The authors themselves rate the 3B version below Suno subjectively, and for Russian it is unusable without transliteration. A 7B version is claimed to match Suno but has not been released.