Voice and avatar tools: the category, and how to judge it

1Why this section names no products

Everything in this category is being rebuilt every few months. Any list of winners printed here would be wrong within two quarters, and worse, it would be wrong in a way you'd trust because it came from us.

So this section teaches the category and the questions. The actual product comparison lives in 05-tool-landscape.md, which is versioned separately and updated on its own cadence, precisely because it goes stale faster than anything else in this guide. Check its last_verified date before you rely on it; if that date is more than 90 days old, it is a list of things to check rather than information.

What follows is the part that doesn't go stale.

2What the category actually contains

Three different jobs get sold under the same banner. Knowing which one you're buying is most of the battle.

Synthetic voice. Text goes in, speech comes out. Two sub-cases that behave very differently: a stock voice the vendor supplies, and a cloned voice built from a sample of a specific person — usually you. Stock voices are instant and legally simple. Cloned voices are the ones worth understanding properly, and the ones with the consent questions attached.

Avatar / talking-head video. A synthetic person delivers your script to camera. Again two sub-cases: a stock presenter from the vendor's library, and a likeness built from footage of a real person. The economics are obvious — one recording session, unlimited scripts — and so are the risks.

Lipsync and dubbing. Existing footage, altered so the mouth matches new audio. Mostly used for translating a piece into other languages without reshooting, and increasingly for fixing a single fluffed line without a pickup day.

Explain-on-first-use: a pickup is a small re-shoot done after the main session to fix or add one specific thing. Traditionally it means getting the talent back. That's the cost these tools are aimed at.

3The seven questions

Ask these of anything in the category, in roughly this order. They are ordered so the ones that can end the conversation come first.

1. Can I use the output commercially, and can I prove it? Not "is it allowed" — where is it written, and can you show a brand the clause? Some tiers permit commercial use and some don't, and the boundary is often the price tier rather than the product. A brand's legal team will ask. Have the answer before they do.

2. Whose voice or face is this, and what did they agree to? Cloning your own voice is straightforward. Cloning anyone else's — a friend, a client's founder, a public figure — is a consent question first and a product question second. Written permission, naming the specific use, is the minimum bar. If you cannot produce that permission, the fact that a tool made it easy is not a defence.

3. Will I be able to get this exact voice back in six months? The one nobody asks and everybody regrets. If you build a client's series on a particular synthetic voice and the vendor retires it, retrains it, or you lose the account, the series breaks. Ask whether voices are versioned, whether you can export or pin one, and what happens to your clone if you stop paying.

4. What happens to the sample I upload? Is it used to train the vendor's models? Can you delete it? Can you get confirmation it's deleted? For your own voice this is a personal call. For a client's founder's voice, it's a question you must be able to answer in writing.

5. Can I hear the tell — and can my audience? Covered below. Judge with your own ears, on your own script, not on the vendor's demo reel. Demo reels are cast to flatter the tool.

6. Does it handle the way people actually speak in my market? Accents outside the vendor's primary training market, code-switching, non-English languages, and names it hasn't seen. This is where tools quietly fail, and it's invisible until it happens to a word that matters — usually a brand name.

7. What does it cost per finished minute, not per month? Sticker price is per month. Your real number is what you spent divided by what you shipped. Some tools are cheap monthly and expensive per output because of render limits or credit systems.

4The tell, in voice

Synthetic speech gives itself away through prosody — the rhythm, stress and pitch movement of a sentence — long before it gives itself away through sound quality. The audio can be flawless and the delivery still wrong.

What to listen for:

• Even stress. Real speech leans hard on one or two words per sentence and throws the rest away. Synthetic speech tends to give every word similar weight.
• Emphasis on the wrong word. I didn't say she stole it has seven different meanings depending on which word you hit. Models frequently pick the wrong one, and it reads as subtly foreign rather than obviously fake.
• Breath in the wrong place, or absent. Real people breathe where the sense breaks. Synthetic breath, when present, is often decorative and evenly spaced.
• No repairs. Humans start sentences again, trail off, say the — the thing. Perfectly clean delivery of casual speech is itself a tell.
• Flat terminal contours. Statements that should fall slightly and don't.

Practical countermeasures, in order of how much they help:

1. Write for the voice, not for the page. Short clauses. One idea per breath. Contractions everywhere.
2. Punctuate for delivery rather than for grammar. A full stop where you want a beat, even if it's a comma splice on paper.
3. Put the emphasis in the script. Most tools respect some form of stress or pacing markup; where they don't, rewriting the sentence so the important word falls naturally at the end usually works.
4. Accept a slightly worse take with better rhythm over a cleaner one that's evenly stressed. Audiences forgive noise. They do not forgive uncanny.

5When not to use this category at all

Worth saying plainly, because the rest of this section is about how to use it well.

• When the brand's whole proposition is a real person. Founder-led brands, small makers, anything where the appeal is that a specific human is standing behind it. A synthetic delivery undercuts the thing they're selling.
• When you'd have to hide it. If a piece only works if the audience doesn't find out, that is not a legal test being failed, it's a taste test. Module 4's standard applies: if you couldn't caption it honestly, don't ship it.
• When your own face is the asset. Plenty of UGC work is booked because a specific creator is recognisably in it. Replacing yourself with a synthetic you can be a rational choice at volume, and it can also quietly remove the reason you were hired.

6Disclosure — confirm, don't assume

Requirements for disclosing synthetic voice and likeness are changing, they differ by platform and by market, and some brands impose stricter rules than any regulator does.

We are deliberately not printing a rule here for you to rely on. Ask the client what they require, check the platform's current guidance yourself, write down the date you checked, and re-check before each campaign. Form BV-01 in Module 3 has a field for exactly this, including the confirmation date, because a rule you verified eight months ago is not a rule you verified.

7Exercise 2.3 — Two voices, one script

Take the hook you shortlisted in Exercise 2.1.

1. Run it through two different tools in this category — any two, including free tiers.
2. Listen to each with your eyes shut, twice.
3. Write down, in your own words, exactly where each one gave itself away. Not "it sounded robotic" — which word, and what was wrong with it.
4. Now rewrite the script to avoid that failure, and run it again.

The output of this exercise is not a voiceover. It's the sentence you can now write, which is the one you'll put in front of a client when they ask whether anyone will be able to tell.