Grom Zvuk

Q: Why is there no streaming, no way to hear part of it earlier?

Grom Zvuk works on every part of the track at once — intro, verses, chorus and ending are computed as one composition. That is why no partial result exists: the track appears when it is complete. In exchange the parts never argue with each other and the mix is consistent across the full length.

Q: Which languages does it sing in?

Any of them — from English to Chinese and Japanese. The model reads the lyrics, the stress and the phrasing of the language, so the vocal sounds natural rather than accented.

Q: How do I make a cover?

Upload the source track and describe the style it should be sung in. The model keeps the melody and the lyrics, and picks the arrangement, instruments and vocal delivery to match the new genre.

Q: Can I get an instrumental only?

Yes. Say in the prompt that no vocals are needed and the model returns an instrumental version at the tempo and mood you asked for.

Q: Is the model available over API?

Yes, once it launches — with the same GPTunneL key as every other model: create a task, wait for the finished track, collect the mp3.

Q: Who owns the rights to the track?

The result of a generation stays yours. For covers of someone else's songs, remember the rights to the original composition — those are governed by copyright law, not by the service.

The first music model in the Grom lineup.

Listen

How Grom Zvuk sounds

Three tracks the model wrote end to end: a rock cover, a pop duet with Passo Bel Solto and a dream-pop piece with Lana Del Rey. Hit a record — it plays right here.

The model launches on 18 September

Technology

How a track is made

First the model floods the whole length of the future song with noise and marks out where the verses go, where the chorus lands and where the bridge sits. Then, over several passes, it turns that noise into sound worth listening to.

IntroVerseChorusVerseChorusBridgeOutro

There is no partial result: the track arrives whole, once it is finished.

Architecture

How the pipeline works

Two experts under shared attention: the first writes music as symbols, the second unfolds those symbols into sound.

CREATEstyle and lyricsCOVERmelody lifted from the sourceEDITedits to a finished score

yPromptstyle · lyrics
sScoreABC notation
cSemantics24 Hz
zAcoustic latents24 Hz × 64
48 kHzFinished recordVAE decoder · stereo

AR · CAUSALPlannercausal token prediction · y → s → c

Walks through the music step by step: decides where the melody leads and where the harmony turns.

NAR · BIDIRECTIONALSynthesiserflow-velocity prediction · z_t → z

Works across the full length at once: refines the sound as a whole, not piece by piece.

HYBRID SELF-ATTENTIONshared memory of the track

The diagram is simplified: it explains the principle rather than repeating the model's internals.

Comparison

Where Grom Zvuk sits among music models

Two axes at once: how the track sounds and how closely the model sings the given lyrics. On lyric accuracy Grom Zvuk is in the top three; on sound it holds its place among the closed services.

sound / lyrics

Suno v588 / 90
Mureka 990 / 78
Grom Zvuk83 / 85
Suno v680 / 85
Suno v5.582 / 82
Suno v4.584 / 78
ACE-Step 1.567 / 74
MiniMax Music 370 / 64
Muse59 / 62
LeVo 265 / 42
DiffRhythm 237 / 53
SongBloom9 / 9

Blind comparison across 150 prompts: identical briefs, finished tracks rated on sound and on how well they follow the given lyrics, with model names hidden from the raters. Values are normalised to a 0–100 scale. Hover a dot to see its name.

Under the hood

What sits behind a generation

48 kHz, stereo

Not a compressed draft but a studio sample rate and a real stereo image.

Tracks up to four minutes

A full song with verses, choruses and an ending, not a fifteen-second loop.

A single pass

Every part of the song is computed together rather than in turn — hence the unified sound.

Our own hardware

The model runs on our own fleet of accelerators, not in someone else's cloud.

Languages

Sings in any language

From English to Chinese and Japanese. The model reads the lyrics, the stress and the phrasing — the vocal never turns into accented mush.

English中文日本語EspañolDeutschFrançais한국어ItalianoPortuguêsTürkçeҚазақшаالعربيةहिन्दीРусский

The list is open-ended: these are not all the languages the model sings in.

Capabilities

Create, cover, edit

Create

Describe the style and the mood, hand over your lines — the model writes the melody, builds the arrangement and sings it.

Cover

Upload a song and name the new sound: the melody stays recognisable while the arrangement changes completely.

Edit

The intent of the track lives in the score, so it can be changed surgically — swap a section or an instrument without rewriting the song.

Need your own stream of music?

Get in touch — we'll go over volumes, limits and pricing ahead of launch. The model runs on our own hardware, so load and timelines are settled by contract rather than by a queue in someone else's cloud.

Message us on Telegram Business solutions