How to Generate an AI Voice: TTS, Cloning, Emotion & More

To generate an AI voice, you type or paste a script, choose a voice and a text-to-speech model, and let the model read it aloud. The difference between a robotic read and a believable performance comes down to three choices: the model, the voice settings, and the small bracketed audio tags (like [whispers] or [laughs]) that tell the model how to say a line. This guide walks through all of it, from your first AI voice to cloning your own voice and putting it inside a video.

Everything below can be done in the Voice tab of the Framiq studio, which runs ElevenLabs models, but the techniques apply to any modern AI voice generator.

What you can make with an AI voice generator

Modern voice AI is no longer just "text to speech". In one place you can:

• Text-to-speech (TTS): narration, voiceovers, explainer videos, audiobook samples, ad reads.
• Voice cloning: a digital copy of your own voice (or a voice you have permission to use) that can read any script.
• Sound effects: footsteps on gravel, a door creaking, a crowd cheering, described in plain words.
• Music: short background tracks generated from a description.
• Dubbing: translate an existing clip into another language while keeping a similar voice.
• Voice changer and voice isolator: turn one recorded performance into another voice, or strip the background noise from a recording.

If your end goal is a video with dialogue, jump to the section on voicing a video. If you just need a clean voiceover, start with the steps below.

How to generate an AI voice, step by step

1. Write the script for the ear, not the eye

AI voices read exactly what you give them. Written text that looks fine on a page often sounds stiff out loud. Before generating:

• Keep sentences short. One idea per sentence.
• Spell out what should be spoken: "twenty-twenty-six", "three point five", "Doctor Martin".
• Use punctuation as direction. A comma is a short breath, a full stop is a beat, an ellipsis is a hesitation.
• Read it aloud yourself once. If you stumble, the AI will too.

2. Choose the model

The model matters more than most people think. Framiq's Voice tab offers four ElevenLabs families:

Model	Best for	Expression controls
Eleven v4	The most natural, expressive reads; dialogue; cloned voices	Audio tags and sound tags (e.g. [door slams])
Eleven v4 Turbo	Fast, cheaper drafts and long narration	Audio tags
Eleven v3	Expressive reads, older projects	Audio tags, style setting
Multilingual v2	Steady, predictable narration in many languages	Style and speed sliders, no tags

A simple rule: start with v4 when emotion matters (ads, characters, storytelling), use v4 Turbo when you are iterating on a long script and want speed, and keep multilingual v2 for calm, consistent narration where you want the same read every time.

One thing to know about v4: it does not use the "style" or "speed" sliders at all. Its entire expression comes from audio tags, which is why the next step matters.

3. Pick or create a voice

Browse the voice library and listen to a few voices with your script, not the sample sentence. A voice that sounds great on "Hello, welcome" may sound wrong on a product description. If you need a voice that belongs to you or your brand, see the cloning section below.

4. Add emotion with audio tags

Audio tags are words in square brackets placed right where the change should happen. They work on v3, v4 and v4 Turbo:

[whispers] I wasn't supposed to tell you this...
[pause] But the launch is tomorrow. [laughs] Act surprised.

[shouts] Everybody out of the pool! [pause] [calmly] Thank you. Now, who brought the inflatable shark?

Common tags that work well:

• Delivery: [whispers], [shouts], [calmly], [excited], [sarcastic]
• Reactions: [laughs], [sighs], [gasps]
• Pacing: [pause], [long pause]

On Eleven v4, you can also place sound tags inline and the model renders them as sound, not speech:

She closed the laptop. [door slams] [long pause] "Who's there?" [whispers] Nobody answered.

Two tips: a tag only works if it suits the voice (a booming, shouted clone will not whisper convincingly), and one or two tags per sentence is plenty. Stacking five tags on a line usually produces something odd.

5. Generate, listen, adjust

Generate, then listen on headphones and phone speakers. If the read is flat, lower the stability setting (lower stability lets the voice move more; higher stability keeps it close to the reference and more monotone) or add a tag. If it is too erratic, raise stability. Regenerate only the line that is wrong instead of the whole script.

How to clone your own voice with AI

Voice cloning creates a voice model from recordings of a real person. Only clone your own voice, or the voice of someone who has explicitly agreed to it. Cloning a celebrity, a colleague or a stranger without consent is a deepfake, it can be illegal, and it is not something we support.

With consent sorted, here is how to get a good clone:

1. Record clean audio. A quiet room, a decent microphone or a recent phone held at a consistent distance, no music, no echo. Eleven v4 can clone from roughly 10 seconds of clean speech; a minute or two of varied speech gives the model more to work with.
2. Speak naturally and with variety. Read a few sentences with different emotions: a question, an excited line, a calm explanation. A clone trained on a monotone recording reads everything in a monotone.
3. Clean up first if needed. If your only recording has background noise, run it through the voice isolator before cloning.
4. Generate on v4. The recording is a property of the voice, not of the model, so an existing clone also benefits from v4's accuracy just by generating on it.
5. Test with tags. Try [whispers], [laughs] and [excited] on your clone to see its range.

If your cloned voice sounds flat, it is rarely the clone's fault: identity comes from the samples, but performance comes from the settings and the tags.

Sound effects, music and dubbing

Sound effects

Describe the sound like a foley artist would: the material, the action, the space.

Heavy wooden door creaking open slowly in a large stone hallway, distant echo

Light rain on a car windshield, wipers on low, muffled city traffic

Vague prompts ("scary sound") give generic results. Specific ones give usable audio.

Music

Give the genre, tempo, mood and instruments, and say what it is for:

Upbeat lo-fi hip hop, 85 BPM, warm Rhodes piano, soft vinyl crackle, background music for a cooking video

Dubbing

Dubbing takes an existing clip and produces it in another language with a similar voice. It works best on clear speech with little overlapping music. For a short social clip, it is often faster than recording a second version.

Voice changer and voice isolator

The voice changer keeps your timing and intonation but swaps the voice, which is handy when you want to act out a line yourself and have it come out in a character's voice. The voice isolator removes background noise from a recording, useful before cloning or before using a field recording in a video.

How to voice an AI video

If you are making AI video, you have two options.

Option 1: voiceover in post. Generate the narration in the Voice tab, then lay it over your video in any editor. Simple and fully controllable.

Option 2: let the video model speak. Seedance 2.5 generates native audio, and you can write dialogue directly in the video prompt. In Framiq, if you link a character to a cloned voice and switch on the Voices toggle, the studio voices each line with that character's cloned voice and hands it to Seedance as reference audio. The rest of the soundtrack (ambience, music, effects) is still generated by the video model. The toggle is off by default because it costs extra.

Write the line with a speaker and a delivery word:

@Lina sits at a café terrace, golden hour, handheld close-up.
@Lina says calmly: {I told you this would happen.}
<cups clinking> (soft jazz in the background)

In Seedance prompts, {…} marks dialogue, (…) marks music and <…> marks sound effects. A line with no identifiable speaker cannot be voiced with a cloned voice, so always name who speaks. Our Seedance 2.5 tutorial and AI video guide go deeper on the video side.

Common mistakes and how to fix them

Problem	Likely cause	Fix
Voice sounds robotic or flat	High stability, no tags	Lower stability, add [excited] or [calmly] where the emotion changes
Tag is read aloud	Model that does not support tags (multilingual v2)	Switch to v4 or v4 Turbo
Style/speed slider does nothing	You are on v4	v4 uses tags for expression; switch to v2 if you need a speed slider
Clone sounds like someone else	Noisy or very short sample	Isolate the voice, record more varied speech
Wrong pronunciation of names or numbers	Written form is ambiguous	Spell it phonetically: "Nguyen" → "Win", "1/2" → "one half"
Odd laughs or noises mid-sentence	Too many tags stacked	Keep one or two tags per sentence

How much does AI voice generation cost?

In Framiq you pay per generation from a prepaid balance. The Free plan uses the standard rate; the Pro plan (30/month) gives you every model at API cost. Text-to-speech is billed by the number of characters, so a short ad read costs a fraction of a cent to a few cents. Exact per-model prices are listed on the pricing page, and failed generations are refunded.

FAQ

What is the most realistic AI voice generator?

The most natural results today come from expressive models like ElevenLabs Eleven v4, especially when you use audio tags to direct the performance. Realism depends as much on the script and the tags as on the model.

Can I clone my own voice for free?

You need a recording of your voice and a tool that supports cloning. On Framiq, a free account comes with free credits to start, and v4 can clone from about 10 seconds of clean speech. Only clone your own voice or one you have permission to use.

Is it legal to use an AI voice?

Using a generated or library voice for your own content is generally fine. Cloning someone else's voice without their consent, or making content that impersonates a real person, can break the law and platform rules. Also note that TikTok and Instagram ask you to label realistic AI-generated content, so use the platform's "AI-generated" label when you publish.

How do I make an AI voice sound emotional?

Use a model that supports audio tags (v3, v4, v4 Turbo) and place tags like [whispers], [laughs], [sighs] or [shouts] exactly where the emotion should change. Lower the stability setting if the read is still too even.

Which AI voice model is best for long narration?

Eleven v4 Turbo is fast and cheaper, which suits long scripts and drafts. For steady, predictable narration in many languages, multilingual v2 is a solid choice.

Can an AI voice speak inside an AI video?

Yes. Seedance 2.5 generates native audio from dialogue written in the prompt, and in Framiq a character linked to a cloned voice can speak with that exact voice when the Voices toggle is on. See our video guide for the full workflow.

Try it

The fastest way to learn is to take one short script, generate it on v4 with no tags, then again with two well-placed tags, and listen to the difference. Open the Voice tab in the Framiq studio, start with the free credits, and when you are happy with the voice, link it to a character and let your videos speak.