Make AI sound like you: voice, Artist DNA and instrument training
Clone your voice from one take, train a voice model for Voice Swap, an Artist DNA on your songs or a model of your instrument. How each works, which plan.
Quick answer: In late October 2026, Soundverse showed four ways to make its AI sound like you: a one-take Voice Clone for singing new songs in your voice, Voice Training for a detailed voice model you use in Voice Swap, an Artist DNA trained on your own songs, and an instrument model you play by humming into Voice to Instrument. They're for singers, producers and artists with a catalogue. Training any of them needs the Creator plan ($12.49 a month) or above, while Voice Swap and Voice to Instrument themselves run on every plan, Free included.
Four ways to train, side by side
| Training | What you give it | Where you use it | Training cost |
|---|---|---|---|
| Voice Clone | One take with 15+ seconds of singing | AI Song Generator, "Use custom voice" | About 6,000 tokens |
| Voice model | 10+ minutes of clean solo voice | Voice Swap | 300 tokens + 30 per epoch |
| Artist DNA | Songs you own (10+ varied songs is best) | The DNA tool | No tokens |
| Instrument model | 10+ minutes of solo instrument recordings | Voice to Instrument | 300 tokens + 30 per epoch |
All four need Creator or above to train, and each one can run from its panel in AI Tools or from a chat with Agent One. Every flow also stops at a consent or rights check before training starts.
One take: Voice Clone
Voice Clone is the fast option. Pick a recording with at least 15 seconds of singing (MP3, WAV or FLAC, up to 10 MB), name the voice, and tick the box confirming it's your voice or that you have the singer's explicit consent. Start Training stays disabled until you tick it.
In the 24 October demo, you record one take and ask Agent One to clone it. Then you turn on "Use custom voice" in the AI Song Generator and pick your clone, and Agent One writes the lyrics, generates the song in your voice, names it and makes the cover art. From there you can open it in AI Studio, separate the stems or turn it into a music video.
Record in a quiet room, keep the vocal dry, and cover a range of notes and emotions. A clone costs about 6,000 tokens to train, more than the 4,000 tokens Creator adds each month. Songs sung in it then cost the same as any other song.
One take, cloned by Agent One, then a full song in that voice.
Voice cloning: a short clip to start, consent first, a deeper model when you have more audio.
Ten minutes of you: Voice Training for Voice Swap
Train Voice Model builds a dedicated model from far more material: at least 10 minutes of clean, dry, solo voice, each recording at least 60 seconds, from your Library or recorded in the panel. Training quality runs from Fast (10 epochs) through Balanced (about 180, the default) to Studio (350), and Soundverse's tutorial calls 150 to 250 epochs typical.
That costs 300 tokens plus 30 per epoch: 600 at Fast, 5,700 at the default and 10,800 at Studio. Training takes hours, with each epoch shown in the chat, and new models start Private.
The finished model shows up under Yours in Voice Swap, which changes the voice but keeps the performance, so the timing and phrasing of the take carry through. Give it a full song and it separates the lead vocal, converts it and remixes it over the original backing. Voice Swap runs on every plan, including Free, at 150 tokens a swap.
Train Voice Model: ten minutes of your voice, trained into a model for Voice Swap.
Train a voice model step by step, then use it in Voice Swap.
Your catalogue as a style: Artist DNA
The 24 October Artist DNA demo starts with one line to Agent One: "Train my artist DNA on these 4 songs." Agent One first asks you to confirm you own or control the rights. Then the tracks upload and train while you watch, an email arrives when the model is ready, and the DNA appears under My DNA, with variants for your next song.
The same training runs from the Train Artist Model panel. One song is enough to start, but 10 or more varied songs give the most faithful DNA. A core-style sensitivity slider (0 to 10, default 5) sets how narrowly it reads your music: lower for a broad genre DNA, higher for narrower style DNAs. Training uses no tokens and usually takes a few minutes.
To make music with it, open the DNA tool, pick your DNA's style model or one of its tracks, and choose a song, an instrumental or a sung vocal. Exporting tracks made with your own DNA is free for you.
Four songs, one request to Agent One, and a trained Artist DNA.
Train Artist Model: pick songs you own, set the sensitivity, confirm your rights.
If you make your DNA public
Publishing is optional. According to the DNA Earn page, you set a public DNA's export price, Soundverse reviews it before it appears in the DNA Marketplace, and 50% of the gross payment for each paid export made with it is reserved for you. Payouts go to a connected Stripe Connect account on a monthly payout date, after a payment-protection period and once you pass a minimum threshold. The page is clear that income isn't guaranteed: you earn only when people pay to export with your DNA or send you Fan Support.
Your instrument, played by humming
Train Instrument Model learns the sound of an instrument you play. Give it at least 10 minutes of solo, dry, close-miked recordings, each at least 60 seconds, covering low to high notes played both soft and loud. It uses the same 10-to-350 epoch slider and the same cost as a voice model, takes hours, and asks you to confirm you own the recordings or are licensed to use them.
When it's ready, the model appears under My DNA and under Yours in Voice to Instrument. In the demo, you hum a line and hear your cello. Voice to Instrument runs on every plan at 150 tokens a conversion, and its built-in models need no training: sing a sitar or bansuri line, or beatbox a rhythm into tabla or dholak.
Train Instrument Model: your own instrument, played with your voice.
Train a model of your instrument, then hum to play it.
How it works: a first weekend with your own models
- Record clean material: a quiet room, the same mic each time, no effects, your full range.
- Clone your voice from one good take and generate a demo in it.
- Train a voice model on 10 minutes of solo vocals, then Voice Swap a guide vocal into your voice.
- Ask Agent One to train your Artist DNA on songs you own, then generate an instrumental from it.
- Train a model of your instrument and hum a new line into Voice to Instrument.
- Pull the results into AI Studio to arrange and mix.
Consent, privacy and rights
Only train on your own voice, music and playing, or on material you have explicit permission to use, and never use a clone to impersonate someone. Soundverse's Settings also has a Privacy tab covering DNA training on your voice, data export and account deletion.
The US Copyright Office has published a report on digital replicas (Part 1, July 2024) and a 2025 report on copyrightability (Part 2), which says prompts alone don't make you the author, while human-authored parts, like lyrics you write and your own edits, can be protected. If you post a video to YouTube, its help page on disclosing GenAI content lists content that "makes a real person appear to say or do something they didn't do". This isn't legal advice.
Which plan
On Free you can already use Voice Swap with Community and Public voices, Voice to Instrument with the built-in instruments, and DNA to make music with other creators' models (exporting those tracks is a purchase at the owner's price). Free-plan music has a "Personal license (non-commercial)".
Creator ($12.49 a month, or $9.99 billed yearly) adds Voice Clone plus voice, instrument and Artist DNA training, 4,000 tokens a month and the Royalty Free licence ("monetize your content"). Pro ($31.24, or $24.99 yearly) gives 10,000 tokens a month and the Sample licence, and Max (from $72, or $60 yearly) carries the Distribution licence, with self-serve distribution coming to Max. Prices are as of November 2026; see pricing and the Terms of Service before you release anything.
FAQ
Should I start with Voice Clone or Voice Training?
Start with Voice Clone if you want new songs in your voice quickly; one take with 15 seconds of singing is enough. Train a voice model when you have 10 minutes of clean recordings and want to re-sing existing vocals in Voice Swap.
Can I clone a famous singer's voice?
Not without their explicit consent. Voice Clone and Train Voice Model both ask you to confirm the voice is yours or that the singer has consented, and training won't start until you do.
Can other people use my voice model or Artist DNA?
Voice models start Private, so only you can use them; a Public one appears for everyone in Voice Swap. Other creators can make music with your Artist DNA once it's public, and Soundverse reviews a public DNA before it appears in the DNA Marketplace.
Will I actually get paid if I publish my DNA?
Only if people use it. Earnings come from paid exports and Fan Support, a refund reverses an earning, and payouts wait for Stripe setup and the minimum threshold. Read the DNA Earn page before you publish.
What recordings make a good instrument model?
Solo, dry, close-miked takes of one instrument, at least 60 seconds each and 10 minutes in total, covering sustained notes, articulations and the full range. Background music or other instruments lower the quality.
For the edit tools these models plug into, read [AI Remix, stems and Audio to MIDI](https://www.soundverse.
Try it: record one clean take and start with Voice Clone, or ask Agent One to train your Artist DNA on songs you own.
