How to Train an AI Voice Model (2026)

To train a voice, use ElevenLabs Professional Voice Cloning on the Creator plan at $22/mo, with at least 30 minutes of your own audio, or the free RVC WebUI with at least 10 minutes of clean speech. An instant clone is faster, and ElevenLabs says it does not train a custom model.

What changed: updated October 2026 with prices re-checked on vendor pages, Kits.ai now in the guide, and Resemble AI and Coqui removed because neither still sells a maintained voice product. A hosted professional clone is the wrong purchase if you cannot record a long take of your own voice, because that plan will not finish the job.

Toolradar data: of the 127 AI voice tools tracked as of October 19, 2026, 90 offer a free tier, and ElevenLabs is listed as an alternative to 52 other tools in that category (AI voice statistics).

An AI voice model is a speech system fitted to one person's recordings so it can say new words in that voice.

A custom model used to require a professional studio, hours of recordings, and a machine learning background. Today it is a recording session plus a paid plan or a free local tool, for video voiceovers, accessibility, or a branded business voice.

For a ranked comparison of the paid cloners, see the best AI voice cloning tools. If you only need a stock voice, how to create an AI voice covers that shorter path.

How much does it cost to train an AI voice?

Prices below were checked in October 2026, and monthly figures are the monthly bill, so a yearly charge is a different commitment from the month beside it.

Sources: ElevenLabs pricing, Kits.ai pricing, Fish Audio plans, RVC's guide, and Speechify's cloning guide.

Skip the instant-clone row when the voice has to hold across a long script, because ElevenLabs says that path does not train a custom model. The professional row is your own speaking voice, Kits.ai is for sung audio or a trained model without a GPU, RVC is the no-subscription row if you will follow a setup guide, and Speechify's API is the row to skip when you need a public cloning price, because it publishes none.

How does training an AI voice actually work?

Training pulls a vocal fingerprint from your recordings: pitch, timbre, tone, rhythm, and cadence. Room noise and a flat read become part of that fingerprint, so a quiet, varied session beats a longer noisy file.

Text-to-speech cloning takes a script and speaks it in the target voice. ElevenLabs and most commercial platforms sell this for voiceover and for text you want read aloud. On ElevenLabs, only Professional Voice Cloning trains a dedicated model. Instant Voice Cloning guesses from a short sample, which is the wrong stop when a listener will hear the voice every week.

Voice conversion, the RVC approach, starts from someone speaking and rebuilds that performance in the target voice, keeping the timing and the emotion. Use it when a performance already exists, not when the only input is a document.

Both still need recordings of the target voice, and the AI Academy teaches the fundamentals behind either approach as hands-on lessons.

How do you train a voice on ElevenLabs?

ElevenLabs is the most popular commercial voice cloning platform, and the hosted path if you do not want to install a training stack. Whether the free plan covers the job is in is ElevenLabs free: cloning does not start on that plan.

Instant Voice Clone

Use an instant clone to hear a likeness today when a rough match is acceptable. ElevenLabs' instant cloning guide recommends about 1 to 2 minutes of clear audio, and says more than 2 to 3 minutes often does not help. The product page also says 1 to 5 minutes can work, so extra length helps only when the clip stays clean. ElevenLabs creates a usable voice clone within minutes of that upload.

Create an account at elevenlabs.io before a serious recording, because the free plan includes 10,000 credits a month and does not include instant voice cloning. In Voices, choose Add a new voice, then Instant Voice Clone, upload a clean sample, name it, confirm you have the right and consent, and test a line in Speech Synthesis.

Instant cloning starts on the Starter plan at $6/mo, with 30,000 credits and a commercial license. For Multilingual v2, one character costs one credit, so that balance is a character budget for that model. The clone captures the general sound but may miss subtle characteristics, especially unique accents or unusual vocal qualities, so do not stop here if listeners will notice those details.

Professional Voice Clone

This is the ElevenLabs option that trains a model, and the one to pay for when the voice carries a series. You need a minimum of 30 minutes of clean, high-quality audio, and closer to 2 to 3 hours for the result the vendor calls best. It has to be your own speaking voice: even with consent, you cannot train a professional clone of someone else, so a client or colleague is out, and singing is out too. Noise, music, a second voice, or a monotone read all train into the model, so record like a conversation.

Choose Professional Voice Clone, upload the files as MP3 at 192 kbps or higher, and verify with a short recording in a similar tone. Fine-tuning usually takes 3 to 6 hours, sometimes up to 24, so test the output before you record another session.

The plan is Creator, at $22/mo after a first month at $11, with 121,000 credits and one professional clone slot. That opening discount is not a train-once exit: if you drop below Creator, the clone stays on the account, but you cannot use it until you are back on Creator or above. Pro is $99/mo with 600,000 credits and the same single slot, so it buys more speech, not a second voice. Scale includes three professional clone slots, and Business includes ten. Paying yearly is priced as 10 months instead of 12, which helps only if you will keep the slot.

ElevenLabs says any microphone can work, and that an XLR mic in the $150 to $300 range, such as an Audio-Technica AT2020 or a Rode NT1, is enough for most voiceover work. Spend that on the professional session, because a phone in a quiet room is enough for an instant clone. Use soft furnishings against echo, stay about 7 to 8 inches from the microphone (about two fists), and read news, stories, and dialogue naturally, because whispering, shouting, or singing will show up in the model.

How do you train a voice with RVC for free?

RVC (Retrieval-Based Voice Conversion) is the open-source path when you will not pay a subscription and you can follow technical instructions. Unlike the text-to-speech tools above, it converts one voice into another while preserving the original speech patterns, emotion, and timing. The project's own guide recommends at least 10 minutes of low-noise speech, and it also documents a local WebUI on Python 3.12 for people who do not want Colab.

You need a Google account for a community Colab notebook, or the project's own WebUI if you would rather install it, plus clean audio of the target speaker. A script alone is not enough, because the model converts a performance.

Collect clean recordings, remove background noise with Audacity (free) or Adobe Podcast's AI noise removal, and export WAV files at 44.1kHz or higher. Noise left in the file becomes part of the voice.

For Colab, search for "RVC v2 Colab", open a community-maintained notebook, connect a GPU runtime, and mount Google Drive so the files survive the session. Those notebooks are maintained by other people, so for the project's own steps, follow the project README and run python webui.py.

Upload the audio and let preprocessing split it into segments and extract pitch, in Colab or in the local WebUI. The project points to RMVPE as the pitch method with the best results, so leave it selected unless you already know why you would change it.

How long training takes depends on the GPU and the number of epochs, and the defaults are the right start. More epochs help until over-training causes artifacts, so stop when a test line sounds right, then feed in a recording and the model converts it toward the target voice. If the quality is off, change the parameters or add clean audio.

The retrieval approach avoids the oversmoothing common in neural models, so the output keeps expression that many text-to-speech clones flatten, which is why the setup is worth it. For other free tools that run voice conversion on your own machine, compare best free AI voice changers. If the voice sits beside images and text, generative AI for content creation shows how those pieces fit.

How do you train a voice on Kits.ai?

Kits.ai is the practical pick when the audio is singing, or when you want a trained model without a GPU of your own. ElevenLabs will not train a sung professional clone, so a vocal catalog starts here. Its help center says Professional Voice Cloning needs 15 to 30 minutes of audio and a few hours to process. Upload 10 to 60 minutes of dry, monophonic vocals, meaning one voice and no instrumental bed, or the mix trains into the model, and training can take 30 minutes to several hours.

The free plan does not include that clone: 0 voice slots and 0 download minutes. Starter is $10/mo for new accounts and includes professional voice cloning, 2 voice slots, 60 conversion minutes, and 15 download minutes. That download allowance is the cap that bites, not training, so you can finish a model and then run out of export time.

Producer is $30/mo with unlimited conversion and voice slots, and 60 download minutes. Professional is $60/mo with unlimited downloads, which is the plan when files have to leave the account. Existing subscribers can be on older allowances, so the price in the account is the one that bills, and canceling freezes the voices until you resubscribe.

Which other tools still belong in the decision?

Fish Audio fits a developer with seconds of permitted audio who will call the voice from an API. Its cloning guide says to use at least 10 seconds, ideally two or three clips of 15 to 20 seconds, and only a voice you have permission to use. A one-off API clone is a reference clip with no training step, and the best sample for that is a clean 10 to 30 seconds. Skip it for a model built on a long performance.

The free tier includes enhanced voice cloning, 8,000 credits a month, and up to 7 minutes of generation, enough to hear the voice and not enough for a series. Plus, the first plan with a professional voice slot, is $15/mo or $132/year. Pro is $100/mo or $900/year and includes 5 professional voice slots. The older S1 model is deprecated and retires on December 31, 2026, so do not start a new workflow on it.

Speechify is a reader for someone who wants text read aloud. The Premium plan is $29/mo on its pricing page, with 1,000+ voices and 60+ languages, for articles, documents, and books. That suits accessibility and personal productivity, not a production clone you ship to other people. Cloning for developers is a separate API: a 10 to 30 second sample plus a spoken consent phrase. That guide does not publish a price for the plan that includes cloning.

Resemble AI no longer sells voice generation: the product page says that offer has stopped, the company now focuses on detecting synthetic media, and the voice models it already built remain free for research, so it is not a place to train a new voice.

Coqui TTS is no longer a service you can sign up for. The company shut down at the end of 2023, and the original repository is no longer maintained, so a new training project should not start from that code.

What should you use a trained voice for?

One training session can feed voiceovers for dozens of videos, which is what pays for a professional clone, while an instant clone is a weak fit because it is not a trained model. People with speech disabilities or degenerative conditions can bank a voice while they still have it and keep communicating in that voice, so record before that take is no longer possible.

A branded voice can cover IVR, automated replies, or training audio without booking an actor for every script change, though ElevenLabs and Kits.ai both keep the voice out of reach if you drop the plan they require. How to use ChatGPT for marketing covers the lines that voice will read. ElevenLabs and similar platforms can generate that voice in 32+ languages while keeping the vocal characteristics, but a sample in one language will carry that language's accent into another.

What consent rules apply to a cloned voice?

Voice cloning can impersonate a real person, so get permission the platform will accept before you upload, and tell the audience when you publish.

ElevenLabs will only train a professional clone of your own voice, and you verify it with a recording. An instant clone asks you to confirm you have the right and consent. A client's voice, even with a release, cannot become an ElevenLabs professional model. Cloning someone without permission is illegal in many jurisdictions and unethical everywhere.

Several U.S. states have passed or are considering laws on voice cloning and deepfakes, so check the rules where you publish. In the EU, Article 50 of the AI Act has required deployers to label deepfakes, including manipulated audio that resembles a real person, since August 2, 2026. A label a person can see or hear is required, and a hidden machine-readable mark is not enough on its own.

Treat the trained model like a password, and do not share model files publicly. ElevenLabs does not let you download a clone as a standalone file, which limits casual leaks and also means you cannot move that voice to another tool. If you need a file you can store yourself, train with RVC.

The AI Academy covers responsible use alongside the practical work.

How do you start training a voice today?

If you are new to this, an instant clone is the shortest path, and it starts on the Starter plan rather than the free plan. Record a short sample on your phone in a quiet room, speaking naturally, upload it, and test the output. You will have a working instant clone in under five minutes. The free plan is still there for stock voices, and it will not clone anyone.

Once that test shows the voice is worth keeping, move to professional cloning for a real model, or to RVC if you want conversion and no subscription. People comparing paid cloners can start with ElevenLabs alternatives.

FAQ

Related guides

• How to Create an AI Voice Model (2026)
• The 8 Best AI Voice Generators in 2026 (Compared and Ranked)
• Best Free AI Voice Generators in 2026
• Browse the best content creation tools

Explore voice cloning, content generation, and other practical AI techniques.

The AI Academy offers courses, tutorials, and hands-on exercises to help you master AI voice model training, audio processing, and content production workflows.

Start your free 7-day trial →