The question we hear most often after an AI agent demo: “Can my own voice call the customers?” Technically, yes, it works and works well. There’s no such button in the AutoCall.kz interface, though, and on the pricing page voice cloning is marked “Limited access.”
This isn’t coyness, and it isn’t a “feature in development.” Below is an honest breakdown: what modern models can do, which of them can be used in a paid service at all, what breaks specifically in Russian and Kazakh, and why a proper voice clone starts with a studio session rather than five seconds of audio.
What the technology can do today
Modern models fall into two classes, and the difference between them is exactly the difference between “wow, sounds similar” and “ready for customers.”
Zero-shot cloning. Models like Chatterbox copy the timbre from a five- or six-second sample, with no training at all. You give it a snippet of audio and it immediately speaks in a similar voice. These are the demos that go viral on social media.
Fine-tuning for a voice. The model is trained on recordings of a specific person: tens of minutes of material and hours of GPU compute. The result is more than a similar timbre. You get the manner of speaking: pauses, typical intonations, the way the person stretches vowels and takes a breath.
You don’t hear the difference in the first phrase. You hear it in the third minute of a conversation. Zero-shot keeps the timbre but is flat in intonation: it doesn’t know how this particular person asks a question or hesitates. In a short notification there’s almost no difference. In a live dialog with an AI agent there is, and it’s noticeable.
Why it isn’t a button in the interface: licenses
Here’s the main reason, which for some reason never makes it into “top 10 AI voice cloning tools” roundups. The best open models are prohibited from commercial use.
Not “you must credit the authors,” not “you must buy a subscription,” but a flat no: the license allows only non-commercial use. Any service that charges money for it is breaking the terms.
| Model | License | Allowed in a paid service? |
|---|---|---|
| XTTS-v2 | Coqui CPML | No. One more detail: Coqui shut down in January 2024, so right now there’s simply no one to buy a commercial license from. |
| F5-TTS | CC-BY-NC-4.0 | No. NC in the name stands for Non-Commercial. Community fine-tunes for Russian inherit the same license. |
| Fish Speech | Fish Audio Research | No. A research license. You need a separate agreement with the rights holder. |
| Chatterbox | MIT | Yes. Zero-shot from ~5 seconds, emotion control, and the Turbo version has latency around 200 ms. |
| Orpheus 3B | Permissive | Yes. Built on Llama-3B, supports zero-shot and controllable emotion. |
| CosyVoice 3 | see release | Yes, with a caveat. It matters because Russian is on the list of supported languages and it has cross-lingual cloning. |
| Qwen3-TTS | Apache 2.0 | Yes. Released in January 2026, StyleTTS2/ISTFTNet architecture, runs faster than real time. |
Licenses change from release to release, and a model that was under Apache yesterday may ship under a research license tomorrow. Before putting anything into production, reread the license for the specific weights version instead of relying on a six-month-old article, this one included.
Hence the “limited access.” Deploying XTTS-v2 on your own servers and charging for it is a quick way to build a service on a legal landmine. We choose from the “Yes” rows of the table, and the choice there is noticeably narrower than the roundups suggest.
What we build it on
We deploy on our own servers, not through someone else’s API. The reason is simple: a customer’s voice is biometric data, and we’re not willing to send it to a third-party service where it would “sit around just in case.” All of our infrastructure is already in Kazakhstan, and for voice prints this is essential.
The baseline choice is Chatterbox under MIT for quick demos and testing a hypothesis, and CosyVoice 3 or Qwen3-TTS for production voices, because they handle Russian and cross-lingual transfer well. The stack is refined for each project: what sounds best on a female voice in a quiet studio doesn’t always win on a male voice with a regional accent.
Hardware: inference needs a GPU. This isn’t a job you can put on a shared application server next to the database. It needs a dedicated GPU node, and that’s a noticeable line in the project budget.
Russian breaks where you least expect it
Say the license is clean, the GPU is bought, and the voice is recorded. Next come the specifics of the language, and for Russian they’re nastier than for English.
Stress. The model has to know which syllable in a word is stressed. English has it easier, since stress there is almost always predictable from the word form. Russian needs a separate preprocessor like RUAccent that places stress marks before the text goes to synthesis.
Homographs. Words that are spelled the same but pronounced differently: zámok (castle) and zamók (lock), ból’shaya (larger) and bol’sháya (big), dórog (expensive) and doróg (of roads). Russian has more than ten thousand of them, and the task still isn’t fully solved even by large neural networks. The Russian National Corpus doesn’t cover every case. A separate unsolved subtask is words where everything depends on the letter yo: vse (all) and vsyo (everything), uznáyem (we will find out) and uznayóm (we recognize).
And it costs time. On a CPU, placing stress marks can take longer than the speech synthesis itself. So the bottleneck turns out to be not the neural network that draws the sound but the boring module that decides where to put the little mark over a letter.
The practical takeaway for the customer: if your text has surnames, company names, place names, or professional terms, you’ll have to set their pronunciation by hand. No model will guess on its own how to read the name of your residential complex or your director’s surname.
Kazakh: almost no data to build on
For Kazakh, the honest word is “shortage.” It is a low-resource language: there is orders of magnitude less training data than for English, and most global models simply don’t support it.
The good news is that the main open corpus was made in Kazakhstan. KazakhTTS2 from ISSAI at Nazarbayev University: 271 hours of recordings, five professional voice actors, a MOS quality score above 4, open data, and a training recipe on ESPnet.
But it’s important to understand that this is a corpus and a recipe, not a ready-made cloning model. You can train good synthesis on it with the voices of those same five speakers. To get a specific customer’s Kazakh voice, the model has to be fine-tuned for that voice separately, which means recording more material than for Russian.
Today in voice broadcasts we cover Kazakh with ready-made professional voices. Cloning a Kazakh voice is a separate project with its own budget, and we say so honestly up front.
How much you really need to record
Online demos promise a clone from five seconds. Formally, they aren’t lying. But between “recognizable” and “the customer didn’t realize a robot was talking” there’s roughly this difference.
- timbre is recognizable
- flat intonation
- no emotion comes through
- “drifts” on long phrases
- good enough to show the idea
- 30–40 minutes of clean audio or more
- quiet room, one microphone
- text marked up with emotions
- questions, refusals, apologies
- names and terms separately
In practice it looks like this: the customer gets a script several pages long with lines marked up by intonation and reads them aloud for the recording. Not “read something,” but a specific set of situations, because the model will learn only what it hears.
— Hello! (friendly, a little faster than usual) This is Example Company calling. — Is now a good time to talk? (question, rising intonation) — I understand. (sympathetic, slower) Then I won’t keep you. — Unfortunately, we had to decline the application. (neutral, without an apologetic tone) — The dove dove into the bushes. (homographs: a control pair)
The last line isn’t a joke. Pairs like this are deliberately included in the recording script so the model hears both pronunciations and doesn’t mix them up later in production.
The flip side: scammers already use this
The same tool that saves a director a day of recording notifications lets someone call his mother in her son’s voice and ask for money. The “relative in trouble” scam existed before neural networks, but the stranger’s voice on the line used to give it away. Now it doesn’t.
A clone doesn’t need much material: a voice message, a story, a recorded talk, a podcast. For a public figure, anyone who wants a sample already has one.
What to do about it in everyday life. Agree with your family on a control question whose answer can’t be googled. If the voice on the phone asks for money or reads out codes, hang up and call back yourself using the number from your contacts. And remember that the number may have call forwarding on, which would send your callback to the wrong person. That’s worth checking separately.
That’s exactly why we don’t make cloning a self-service feature. A button that says “upload any voice file and start calling” isn’t a product. It’s a fraud tool that would bring more problems than money within a week.
How to get access with us
Voice cloning at AutoCall.kz is available on individual request, per project. Here’s how it works:
- Request and task. Who we’re calling, what we’re saying, and what volume. This decides whether you need a clone at all or a ready-made professional voice is enough. Sometimes it is, and we’ll tell you so directly.
- Confirming rights to the voice. The owner of the voice gives written consent. If the voice belongs not to the customer but to, say, an employee or a hired voice actor, the consent must come from that person personally. This isn’t a formality. It’s a condition of working with us.
- Recording the material. A marked-up script, a recording session, an audio quality check.
- Training and acceptance. We listen to the result on real texts from your campaigns, not on demo phrases.
- Launch. The voice is available only in your account.
Timing and cost are calculated for each task: they depend on the language, the amount of material, and whether your project needs a dedicated GPU node. Kazakh costs more than Russian, for the reasons described above.
If the task is simpler than it seems, for example you just need a pleasant human voice rather than your own, start with regular speech synthesis: it’s available right away, costs next to nothing, and in half the cases solves the problem completely.