What are RVC voice models, and can you use one to sing your own song?

You have a finished lyric in one window and two files in another: alto_v2.pth and added_IVF512_alto_v2.index. Somebody on a forum called them "a voice". You open the RVC interface, look for the box where the words go, and there is no such box. There is a slot for an audio file, a pitch number, a dropdown full of algorithm names, and nothing that accepts text.

That empty slot is the whole answer to the question in the title, so it is worth understanding before you download anything else. An RVC model is not a singer. It is a costume. Somebody still has to sing the song, and the model changes what that singer sounds like.

This page is for the person who wrote (or is about to write) their own song and wants to know whether an RVC voice can perform it. If what you actually want is to put a famous voice on a song that already exists, that is a different job with a different legal shape, and the guide to AI cover song generators is the page for you. Here we stay with your own lyric, your own melody and a voice you are allowed to use.

What RVC is, in plain terms

RVC stands for Retrieval-based Voice Conversion. The original project, Retrieval-based-Voice-Conversion-WebUI on GitHub, is open source under the MIT licence and describes itself as a voice timbre conversion and voice changer framework. Several forks build on it (Applio is one), and they share the same basic idea.

The idea works like this. You feed in a vocal recording. The software pulls out what was sung (the pitch line, the timing, the shape of the words) and separates it from who sang it. Then it rebuilds the sound using the timbre of the voice model you loaded. The "retrieval" part is the clever bit: the project's own README says it replaces source features with features from the training set using top-1 retrieval, to cut down on the original singer's tone leaking through.

What that means in practice:

Read the right-hand column twice. The model contributes a sound and nothing else. It does not write a melody, it does not know your lyric, and it will faithfully reproduce a flat note in its own beautiful timbre.

A few facts from the project page help calibrate expectations. The README recommends at least ten minutes of low-noise speech to train a model, which is a small amount of data by machine learning standards and is why so many voice models exist. The pretrained base model, it says, was trained on nearly 50 hours of audio from the open VCTK dataset. And the project includes a real-time voice changer with a reported end-to-end latency of 170 ms, or 90 ms with ASIO audio devices. That real-time mode is how streamers use RVC; for a song you will almost always convert a recorded file instead, because you want to choose the best take first.

Where voice models come from

A trained RVC voice usually travels as two files. The .pth file holds the model itself. The .index file is an optional companion that the retrieval step searches; with it, conversions tend to sound closer to the training voice, and without it they still work.

People get these files in three ways.

They train their own. Ten or more minutes of clean, dry recordings of one voice, run through the training tab of RVC or a fork. Training wants a GPU; the RVC README claims it is fast even on weak graphics cards, and there is a CPU-only install too, which you should expect to run slower.

They download a shared one. Hugging Face hosts a large number of community RVC models, and dedicated model sites exist too. The quality ranges from excellent to unusable, and the paperwork ranges just as widely: some model cards say research use only, many say nothing at all.

They use a hosted service that wraps voice conversion in a website, with its own licensed voices and its own training button. More on those below.

Here is the uncomfortable part about the second route. A large share of the popular shared models are trained on famous singers, from recordings nobody licensed, with no consent from the person whose voice it is. The file being free to download tells you nothing about whether you may publish a song sung in that voice. The MIT licence on the RVC code covers the code. It does not cover the voice inside a model somebody else trained.

Can you sing your own song with one? Yes, with a guide vocal

Here is the full pipeline, start to finish. Every arrow is a separate step with its own tool.

lyric.txt          your words with [Verse] / [Chorus] tags
   |
guide_dry.wav      a person (or an AI song's vocal stem) singing that lyric, no reverb
   |
voice.pth + .index a voice model you have the right to use
   |
converted.wav      RVC output: the guide's melody and timing, the model's timbre
   |
final mix          converted vocal over your instrumental

The two steps people skip are the first and the second. They get excited about the third.

Step 1: the lyric

RVC does not care about lyric structure, but the guide singer does, and so does any AI song tool you use to build the guide. A lyric with clear sections, short singable lines and a chorus that repeats is easier to perform in one clean pass.

This is the step our AI song generator handles. You fill a form (point of view, language, structure, hook type, vocal gender, emotion, listener age, and a theme of up to 800 characters) and it returns a title and a full tagged lyric ready to paste into Suno's Lyrics field. For example, a brief like this:

POV        first
language   en
structure  classic
hook       title-line
vocal      female
emotion    longing
age        25-34
theme      She still leaves the porch light on for someone who moved away.
           His coat is in the hall, the letters she never sent are in a drawer,
           she pours two cups of tea out of habit.

comes back as a lyric shaped like this (an excerpt):

Porch Light Left On

[Verse 1]
Your coat still hangs where the hallway bends
A drawer of letters I never sent
The kettle clicks and I pour for two
Then stand there holding the cup for you

[Chorus]
Porch light left on, porch light left on
Burning for someone already gone
I tell myself it is only a light
Porch light left on through the night

Each image in it traces back to the theme line: the coat, the letters, the second cup, the light. That matters for a guide take, because a lyric with concrete pictures gives the singer something to act, and the acting is exactly what RVC carries over.

One generation on the SM1 model costs 20 credits, which is $0.20. The site is independent: it is not Suno, and it is not connected to the RVC project, Kits AI or any other company named here. It writes words. It does not sing, it does not output audio, and it does not clone voices.

Step 2: the guide vocal

This is where the real performance gets decided, so give it the most attention. You have two honest options.

Sing it yourself. Even if you hate your voice. RVC will replace the timbre; what it keeps is your phrasing, your timing and your pitch. If you can sing in tune and roughly in the right range, your own take is the best guide there is, because every choice in it is yours. Record dry: no reverb, no delay, no doubling, one voice, a quiet room.

Generate the guide. Paste the lyric into Suno, write a style prompt, and generate until a vocal performance has the phrasing you want. Then export the vocal stem and use that as the guide. The melody now comes from the AI song, and RVC re-voices it. Before you publish anything built this way, check what Suno's terms say about using output on your plan; our piece on who owns AI-generated music walks through the questions to ask.

If you go the generated route, a few Suno habits make a better guide: name a single lead vocal in the style prompt, avoid choirs and stacked harmonies in the parts you want to convert, and keep the arrangement sparse so stem separation has less to untangle. Our guide to locking a vocal persona covers keeping the same singer across sections, which helps when you convert the whole song in one pass.

Step 3: the conversion

Load the model, load the dry guide, set a handful of controls, convert. Labels differ between forks, but most RVC front ends expose roughly these:

Convert a single chorus first and listen before running the full song. Range problems show up in the highest line, and a chorus is usually where that line lives.

Step 4: the mix

The converted vocal is a dry track. Treat it like any other lead vocal: level it, add the reverb and compression you left off the guide, sit it over the instrumental. If you generated the guide in Suno, the instrumental stem from the same song is the obvious backing.

What RVC cannot fix

The guide vocal is the ceiling. Everything below follows from that one fact.

Pitch. A sharp note in the guide is a sharp note in the output. Some front ends will let you correct pitch, but RVC as a converter does not.

Diction. Pronunciation comes from the guide. If the guide singer swallows the ends of lines, so does the converted voice.

Range. A model trained on a voice that never sang above a certain note has little to retrieve from up there. Pushing a guide far outside the training range is where the metallic, warbling artefacts appear. Transposing the guide into the model's comfortable range is usually better than forcing it.

Mess in the input. Reverb, backing vocals, a guitar bleeding into the mic: the converter tries to voice all of it. A dirty guide gives a dirty output, which is why the word "dry" keeps coming back on this page.

Emotion that was never there. If the guide is a flat read of the lyric, the output is a flat read in a better voice. This is the most common disappointment, and it has nothing to do with the model.

Whose voice it is: consent first

Three cases, from clear to murky.

Your own voice. Train on your own recordings and you are converting yourself into yourself, which sounds pointless until you realise it lets a better singer (or a Suno vocal) perform your song in your timbre. This is the cleanest use there is.

A collaborator's voice. Ask, in writing, and agree what the model may be used for and who keeps the files. A model of someone's voice is easy to copy and hard to recall.

A famous singer's voice. This is where the law is moving fastest. Tennessee's ELVIS Act, signed in 2024, explicitly protects a person's voice including digital imitations. At federal level, the NO FAKES Act on digital replicas has been introduced in Congress several times, most recently in 2026, and was not law at the time of writing. Our page on celebrity voice generators and the law goes through it properly. Outside the US the rules differ by country, and in some countries personality rights cover the voice too.

This is a description of the landscape, not legal advice. Our disclaimer sets out what we do and do not claim about the lyrics our generator writes.

Three ways to get your song sung in a voice that is not your raw take

RVC is not the only route, and for some readers it is not the best one.

A note on each, with the facts as stated on the companies' own pages at the time of writing.

Kits AI describes a library of more than 100 royalty-free artist voices, which it says are licensed from the artists themselves, plus voice creation and instant voice cloning for a custom singing voice. It has a free tier and paid plans that start at $10 a month at the time of writing. If you are deciding whether it is worth paying for, our Kits AI review goes deeper.

Suno Voices is a Suno feature that lets you add your own voice to Suno-generated songs. You supply 15 seconds to 4 minutes of audio, pick the best two minutes, and then read a short phrase aloud so Suno can check that the speaking voice and the singing voice belong to the same person. Suno's help page says this is there to stop people using the voices of others (it names collaborators, creative partners and public figures) without permission, that you must confirm you hold the rights, and that the feature is for users 18 and over and may not be available everywhere. Plan availability is not stated on that page, so check your plan. The practical difference from RVC: Suno generates the song and sings it in your voice instead of a default singer, so you skip the guide step entirely, but you also give up control over the melody.

If your only goal is hearing your own lyric in your own voice, Suno Voices is the shorter road. If you want a specific melody you wrote, or a voice that is not yours (with permission), RVC or a hosted converter is the tool, and the guide vocal is where you keep control.

Where the lyric step fits, honestly

Every route above starts with words. RVC needs a guide singer to sing them; Kits needs a vocal to convert; Suno Voices needs a lyric in the Lyrics field. None of them writes the song for you in a form you would be happy to perform.

That is the narrow job the generator does: a title and a full lyric with section tags, built from your brief, in one of nine languages. SM1 costs 20 credits ($0.20); SM2 and SM3 cost 80 credits ($0.80) and run an extra editing pass, where a detector flags weak lines and a repair step rewrites them. Credits come from a plan: $5 a month for 500 credits, or $19 once for 1,500 credits that never expire. New accounts start at zero, so there is no free trial of the generator. The details are on the pricing page.

Two limits so nothing gets oversold. Lyrics from the generator are not exclusive and can overlap between users. And nothing we sell sings, converts or clones a voice; that part belongs to whichever tool you pick in step three.

Your first evening with RVC, as a checklist

1. Decide whose voice it is, and get written permission if it is not yours.
2. Get the lyric finished first: tagged sections, short lines, a chorus you can sing twice without running out of breath.
3. Record or generate one dry guide take of the chorus only.
4. Convert the chorus with RMVPE and the index loaded. Adjust transpose until the highest line stops warbling.
5. Only when the chorus sounds right, record or generate the full guide and convert the whole song.
6. Mix the converted vocal like a real one, then listen on headphones and on a phone speaker before you call it done.

If step 2 is where you stall because the words are not there yet, start with the lyric and come back to the voice when you have something worth singing.