How to make an AI voiceover: natural text to speech, step by step

This guide is for YouTubers, course creators, podcasters and small businesses who need a voiceover but have no microphone or time for retakes. It shows how to make an AI voiceover in Fattly: which model to pick, how to write a script that sounds human, what it costs and when to clone your own voice.

What you get

• An MP3 file of your text read aloud. One recording holds up to 1,000 characters, roughly a minute of calm speech.
• Six text-to-speech models that read English. Three of them let you choose from 28 voices, female and male.
• Optionally, a recording in your own voice, cloned from a short sample.

How much an AI voiceover costs

In Fattly you pay credits per recording, not per character, so a short sentence costs the same as a full 1,000 characters. You see each model's price in the list and the cost of the recording on the button before you generate anything.

You get 10 free credits at sign-up, no card required. For voiceovers that goes a fair way: three recordings on Kokoro (3 credits each), one on ElevenLabs v3 (9 credits), or a side-by-side test of Gemini TTS (6) and Chatterbox (4) on the same short text. It won't cover voice cloning (150 credits). The smallest pack is 150 credits: 50 Kokoro recordings, 25 on Gemini TTS, 16 on ElevenLabs or MiniMax, or one recording in your cloned voice. Current pack prices are on the pricing page.

1. Open the voice generator and choose the Voiceover tab

Log in (you need an account with a confirmed email address) and pick Voiceover from the Audio menu. The tool has three tabs: Voiceover, My voice and Music. The voice language follows the site language: on the English site, Fattly tells the model your text is English (US English in Gemini TTS) instead of letting it guess. For a voiceover in Polish, German, Spanish or French, switch the site language first. Kokoro speaks English only.

2. Paste a script written for the ear

Paste your script into the “Text to read” box. The counter underneath shows how many characters you have used: the limit is 1,000, or 300 in Chatterbox. Split longer material, such as a course lesson, into parts under 1,000 characters. Cut at the end of a paragraph, not mid-sentence, so the joins sound natural in your edit.

3. Choose a model and a voice

Pick a model in the “Voice engine” field, where each one shows its price, then choose a voice from the “Voice” list. For ElevenLabs voices, the “Listen” button plays a sample before you spend any credits. Chatterbox and Kokoro have no voice list. The tool opens with ElevenLabs v3 selected. Where to start with English text:

• Kokoro (3 credits): the cheapest, with one fixed voice. Good for checking a script's rhythm before paying for a premium voice.
• Gemini TTS (6 credits): 14 voices, almost all female, with natural intonation and support for pause markers.
• MiniMax Speech 2.8 HD (9 credits): male voices (Casual Guy, Deep Voice Man) at the same price as ElevenLabs, plus five female voices.
• ElevenLabs v3 (9 credits): the most expressive, with emotion tags. For ads and anything where the voice has to act.
• ElevenLabs Multilingual v2 (9 credits): a natural narrator with the same 7 voices as v3.
• Chatterbox Multilingual (4 credits): an open model for short lines of up to 300 characters, no voice choice.

The easiest way to decide: generate the same paragraph in two models and keep the better one. A test like that costs a few credits.

4. Generate, listen and download the MP3

Click “Generate voice”. The button shows the price, for example “Generate voice · 9 cr”. The recording is usually ready in a few seconds and starts playing straight away. If something sounds off, change the text (different punctuation or a number written out in words usually fixes it) and generate again. The download is an MP3 with “ai-generated” in its name, and it stays in your history for 30 days. If a generation fails for technical reasons, your credits come back.

How to write a script that sounds natural in text to speech

The model reads exactly what you write, punctuation included. Most problems with how a voiceover sounds are fixed in the script itself:

• Short sentences. One sentence, one idea, ideally 15 to 20 words at most. Long, nested sentences come out flat.
• Punctuation is pacing. A period gives a clear pause, a comma a short one, an ellipsis lets the voice trail off, and a question mark lifts the intonation. If the voice rushes, add periods.
• Numbers, prices and dates in words. Write “thirty percent” instead of “30%” and “March fourth” instead of “3/4”, which could also be read as “three quarters”. Write a price the way you would say it: “nineteen ninety-nine”.
• No abbreviations or symbols. “For example” instead of “e.g.”, “and” instead of “&”. If a voice mangles a brand name, spell it the way it sounds.
• Read it out loud. If you stumble over a line, the AI voice will sound awkward there too.

Two models also understand markers in square brackets. Gemini TTS responds to markers such as [short pause] and [excited], and ElevenLabs v3 has emotion tags, for example [excited]. Put the marker right before the line it should affect. Don't use markers in the other models or with your own voice, because the voice may read them out loud.

4 voiceover scripts to copy

Each one fits in a single recording, with the numbers already written out. Swap in your own names and details.

• YouTube intro: “Got half a zucchini, two eggs and some leftover rice in the fridge? That's dinner sorted. Today I'm showing you three recipes made from leftovers. Each one takes under twenty minutes, and the last one… genuinely surprised me. Let's get into it.”
• Product ad: “No time for breakfast? Daybreak Oats is ready in three minutes. Add hot water, stir and eat. No added sugar, just real pieces of fruit. A box of seven breakfasts is twelve ninety-nine. Order today and we'll ship it tomorrow morning.”
• E-learning or tutorial: “Lesson three: how to add a new client. Step one. In the menu on the left, click Clients. Step two. In the top right corner, click Add client. Step three. Fill in the first name, last name and email address. Fields marked with an asterisk are required. Finally, click Save. That's it. In the next lesson, I'll show you how to send your first invoice.”
• Podcast intro: “Hi, I'm Martha, and this is The Back Office, a podcast for people who run a small business and still want a life outside it. In episode forty-two, I talk to the owner of a neighborhood bakery about how she raised her prices without losing customers. Grab a coffee. Let's begin.”

A podcast intro sounds best in your own voice, and a YouTube intro usually needs a music bed. You can make one from a description in the Music tab: 15 seconds costs 12 credits, 30 seconds 20, and a full minute 35 (default model Stable Audio 3.0).

My voice: AI voice cloning from a short sample

If the voiceover should sound like you, use the My voice tab. You upload a sample of your voice, type the text, and the model reads it in the voice from the sample:

• Record 10 to 30 seconds of clear speech: a quiet room, no background music, one speaker. Speak in the tone you want in the final recording. A phone is enough, and MP3, WAV, OGG, M4A and AAC files all work.
• Type the text you want spoken, up to 1,000 characters.
• To keep using this voice, give it a name in “Save this voice for reuse (optional)”.
• Tick the declaration that this is your own voice or that you have the consent of the person speaking in the sample. You can't generate without it.
• Click “Speak in my voice”. Cost: 150 credits.

After that, a saved voice costs much less. Pick it from the “Voice” list in the My voice tab and each new recording costs 40 credits instead of 150, with no new sample. It also works in Ad Studio for a “Hands + voiceover” ad (150 credits for the video plus 40 for your voice) and in Dubbing, which translates a clip of up to 10 seconds into another language (35 credits). To be upfront: the model provider stores the saved voice, and it can expire if unused for about a week, so the list only shows voices used in the last 7 days. If yours is gone, create it again from the same sample.

Celebrity and character voices: what's allowed

You can clone only your own voice or that of someone who has expressly agreed, including to how the recording will be used. Our Terms prohibit impersonating other people, so a celebrity's, politician's or influencer's voice without consent is off the table, even as a joke. Fattly has no celebrity voices and no voices of well-known cartoon or game characters. For your own original characters, like the narrator of a children's story or a cartoon villain, pick a fitting voice from the list and direct it with punctuation and, in ElevenLabs v3, emotion tags. If you publish a recording that realistically imitates a real person, you must clearly disclose that it was made with AI.

Common mistakes

• Digits and symbols. A line like “3 x 2” can be read several ways. Write it out in words.
• Too many short recordings. You pay per recording, not per character, so one 900-character paragraph costs less than five separate sentences.
• Different voices in one project. Stick to one model and voice per video, or the parts will sound like different people.
• The wrong site language. On the English site the model is told your text is English, so a Spanish script can come out with an English accent.
• A poor cloning sample. Noise, echo, music or a second speaker all carry over into the clone.

Summary

An AI voiceover now comes down to a script, a model and a few seconds of waiting. In Fattly a recording costs 3 to 9 credits, and the 10 free credits cover your first ones. Keep sentences short, write numbers out in words, compare two models on your own script, and if the voiceover should sound like you, clone your own voice from a sample.