Two Clones, One Render

A video crossed my feed where the person teaching you how to clone yourself... is the clone. Face, hands, voice, the little head tilt- the whole ten minutes is an AI avatar walking you through how it was made. That's a pretty good sales pitch for the method, I'll give it that.

Then the very first "must have" in it turns out not to exist anymore. So this is me saving the method, fixing the numbers that went stale, and looking at what else does the same job for less... or for free. Mostly for me to remember, but maybe it saves you a subscription.

The Method, Two Clones Glued Together

Two clones. A video clone of your face and body from HeyGen (they call it a Digital Twin), and a voice clone from ElevenLabs. HeyGen's own voice is "okay" per the video; the ElevenLabs one is what sells the illusion. You glue them together with an API key so HeyGen speaks with your ElevenLabs voice.

What's the part everyone gets wrong? The training footage- garbage in, uncanny valley out.

Filming Your Face- Five Outfits, One Session

Five clips, each three to five minutes long, a different outfit in each one. Different shirt so your avatar isn't wearing the same thing in every video... people notice that.

Everything else stays the same. Same camera, same mic, same lighting, same session. I pulled HeyGen's own spec for the footage so I'm not guessing later:

• At least 1080p at 30fps. 4K/60 is better.
• One continuous take. No cuts, no splicing. Flub a line? Keep talking or restart the whole clip.
• Talk the whole time- the audio in the clip is what the lip-sync learns from.
• Bright and even, no harsh shadows, plain static background, quiet room with no other voices.
• Gestures stay in frame, below the chest mostly, never over your face.
• Two minutes is the minimum per clip, up to five recommended.

The tip I'm keeping from the video, the one nobody writes into a spec sheet- film when you're energized. The presenter filmed a take while exhausted and the clone came out looking... tired. Forever. Your mood is training data too.

Your Voice Takes Way More Tape

The voice clone wants a lot more material than the face does. The video says 60 to 90 minutes; ElevenLabs' own guide says 30 minutes is the floor, two to three hours is what they recommend, and 180 minutes is the cap. So 60-90 works, it's just not the ceiling.

Rules for that audio- one speaker only, no music, no background noise, no room echo, same mic the whole way, and cut the long pauses and the um's. MP3 at 192kbps or higher; they actually say WAV is not recommended, which surprised me.

The trick in the video was pulling ten old videos that were just the presenter talking and stitching them into one file. If you've got a backlog of recordings, that's your dataset. If not... read a book out loud for an hour. Here's the stitch with ffmpeg. It evens out the loudness too. Name the files in order first (clip01, clip02... the glob sorts clip10 before clip2), keep apostrophes out of the filenames, and run it in a folder with only those files in it:

mkdir -p mp3
for f in *.mp4; do ffmpeg -i "$f" -vn -ar 44100 -b:a 192k "mp3/${f%.mp4}.mp3"; done
for f in mp3/*.mp3; do echo "file '$PWD/$f'"; done > list.txt
ffmpeg -f concat -safe 0 -i list.txt -af loudnorm=I=-20:TP=-1.5:LRA=11 -ar 44100 -b:a 192k voice-training.mp3

In ElevenLabs it's Voices, Create Voice, Professional Voice Clone. Upload, describe the voice, name it. Only turn on background-noise removal if the mic was bad- cleaning audio that's already clean just costs you fidelity.

Both of Them Want to See You, Live

You can't clone someone off a YouTube rip. That's on purpose- honestly, I'm glad it is. There are two separate checks here, one per service. ElevenLabs makes the person being cloned read some lines off the screen, live, before training starts.

HeyGen wants a consent video under 30 seconds, read word for word from the script it shows you- sometimes with a code in it, so it can't be pre-recorded. It has to be the same person as the footage, recorded in the same session, and not a screen recording.

Doing this for a client who's somewhere else? The workaround from the video- get them on a video call with a decent mic and have them read the lines off your screen through your device.

Wiring the Voice Into HeyGen

In ElevenLabs, make an API key (it lives under your profile, API Keys). When you create it, either turn off Restrict Key or give it these permissions, otherwise HeyGen throws "Invalid API key" at you:

Text to Speech, Voices, Voice Generation, Models, History, User

Over in HeyGen you start a video the normal way (Create, then Create a Video, then Create in AI Studio) and pick your avatar. Hover the little person icon on the scene and the Voice dropdown shows up... from there it's + New Voice, then Integrate 3rd Party Voice, then Import Voice. Pick ElevenLabs, paste the key, Confirm. Those are the menu names as of September 2026- HeyGen moves things around, so if a label's changed, their help article (in the sources) gets updated faster than I will.

One gotcha HeyGen spells out itself- the voice runs on your ElevenLabs credits. Run those dry and your HeyGen renders using that voice stop working, so there are two meters ticking.

Writing for a Robot Mouth

You can't paste any old script in. TTS trips over things a human reads without thinking... numbers get written out as words ("50%" becomes "fifty percent"), every Enter is a pause so each line break gives it a breath, and emojis and special characters go- nothing it has to guess how to say.

The prompt from the video is the bit I'll actually reuse- it works in any chatbot:

Rewrite this for AI voice reading. Convert all numbers to words. Remove filler. Make it clear and natural.

Chatbots miss stuff though, so read what comes back and spell out any stray %, $ or & by hand. When it's clean, paste it into HeyGen and make sure the ElevenLabs voice is the one picked in the dropdown. Clicking at the top of the timeline makes the avatar fill the whole thing. The video sets 30 FPS and 1080p before hitting Submit.

Don't publish the raw render. B-roll, music, text on screen, transitions... that's what gets it from "huh, that's an avatar" to "wait, was that an avatar?" The video's answer is a paid editor; DaVinci Resolve is free if you're doing it yourself.

The Plan in the Video Is Gone

Here's where it falls apart. The video says the one you want is the HeyGen Team plan at $89 a month- 30 minutes of video and, "most importantly", custom video avatars, which it calls non-negotiable. It says the $29 Creator plan only does instant avatars.

HeyGen's pricing page, as of September 2026- there is no Team plan. The lineup is Free, Creator $29, Pro $49, Business $149 plus $20 a seat, and Enterprise. Custom video avatars aren't locked behind anything anymore- even the Free plan gets one Digital Twin slot. Creator and Pro get five by default; Business gets ten (some new Business signups since August 6, 2026 get five).

So the $111 a month in the video ($89 + $22 for ElevenLabs) is really $51 today ($29 + $22) for the same setup. Or $29 if you skip ElevenLabs, since Creator now lists its own voice cloning.

But- and this is the salt in the burger- HeyGen doesn't sell minutes anymore, it sells credits. Their own Avatar IV guide says a video-look avatar (the Digital Twin kind) burns 31 credits a minute; a photo look burns 16. My napkin math, at today's rates- Creator's 600 credits is about 19 minutes of Digital Twin video a month. Pro's 1,000 is about 32. Business, 1,500 credits, about 48.

So Creator isn't "30 minutes" of your twin. If you're doing three videos a week at four minutes each, that's around 52 minutes a month... you're in Business territory, or buying top-ups. Count your minutes before you pick a plan.

Other timing claims in the video that the docs don't back up- voice training "about two hours" (ElevenLabs says roughly three to six, longer when busy), avatar processing "5 to 10 minutes" (HeyGen says around 10 to 20 for non-4K)... I'd plan for a day and be pleasantly surprised.

Does Anyone Do It Better for Less?

Depends which half. For the face, the honest answer is not really- for the voice, yes, a few do.

The face. Synthesia is the big other name. Its Creator plan is $89 a month ($64 billed yearly) for 30 minutes and five personal avatars; the $29 Starter gets three personal avatars but only 10 minutes of video a month. Against HeyGen Pro at $49 for about 32 minutes of twin, HeyGen's cheaper per minute. Watch for tools that sell the plan cheap and the clone separately- Elai lists its selfie avatar as a $199-a-year add-on and voice cloning as another $200 a year, on top of the subscription. Tavus starts custom replicas at $59 a month, but that's aimed at live conversational video agents, a different job.

The voice. This is where the money is. All prices off their own pricing pages:

• HeyGen's built-in clone- included from Creator up. Free is free... test it against ElevenLabs before paying for both.
• Hume Creator, $14/mo ($7 the first month)- 140,000 characters, voice cloning listed as unlimited.
• Cartesia Pro, $5/mo- 100K credits with instant voice cloning. Their higher-fidelity Pro Voice Cloning starts on the $49 Startup plan.
• Fish Audio Plus, $11/mo- up to 200 minutes of generation and one professional voice slot.
• ElevenLabs Creator, $22/mo- the one in the video. The video plays it side by side against HeyGen's stock voice so you can judge for yourself. Starter at $6 only gets instant clones, no Professional Voice Clone.

Instant clones work off seconds of audio instead of hours- quicker but usually less "you". Whether $5-$11 gets close enough to ElevenLabs' professional clone for your voice... that's a ten-minute test a pricing page can't answer. Most of these have a free tier, so it costs nothing to find out!

The Free Route, If You've Got the GPU

Voice is solved locally. Chatterbox from Resemble AI is MIT licensed, clones from a short reference clip (their examples use ten seconds) and runs on CUDA, Apple Silicon (MPS) or plain CPU. Heads up- every file it makes carries Resemble's Perth watermark, inaudible but detectable. Feature or bug depends on who's asking.

Face is... not solved. The open models that lip-sync a real person to new audio:

• MuseTalk. MIT, fine for commercial use, real-time (30fps+ on a Tesla V100), video or image plus audio in. But the face region is 256 x 256 pixels. On a 1080p video, that's the tell.
• InfiniteTalk. Apache 2.0, image or video plus audio, long-form, 480p and 720p. Wants a real NVIDIA card; there's a quantized model and a low-VRAM flag if yours is smaller. The team behind it shipped LongCat-Video-Avatar-1.5 in May 2026, so it's alive.
• LivePortrait. Runs on Apple Silicon (humans mode), but it's driven by a video of you moving, not by audio. More puppet than talking head.

None of them get you a HeyGen-grade twin at 1080p that holds up for ten minutes. For a short clip, a meme, a talking-photo bit? Sure. For a channel? Not yet- and I'd love to be proven wrong on that one.

Nobody Noticed... Should They Have?

The video brags that nobody in the comments asked if the first avatar video was AI. That's the pitch. YouTube's rules make it a bit more interesting. Its disclosure page says this doesn't need a label:

Cloning one's own voice to create voice overs or dubs

...but that it does need one when content makes "a real person appear to say or do something they didn't do". Your avatar reading a script you wrote- is that you saying it, or a real person appearing to say something they didn't? Grey. The penalty for creators who "consistently choose not to disclose" can go as far as removal or suspension from the Partner Program. Personally I'd tick the altered-content box and let the B-roll do the impressing.

That's the whole rig... it's way cheaper than the video makes it sound, if you count your minutes first. Enjoy.

Oh, the video's discount link still redirects to HeyGen with a referral tag on it- it's an affiliate link, just so you know what you're clicking. I didn't sign up to see if the 20% code still takes.

Anyone cloned themselves already? Which voice did you land on- and did anybody in your comments clock it? Leave a comment, I want to hear how close the cheap ones get.

Related on here- the AI that calls you, a voice agent setup, and speech to text, fast, easy, free for going the other direction.

Glossary

• Digital Twin — HeyGen's name for a video avatar trained on footage of a real person.
• Look — one outfit, background or angle of the same avatar; HeyGen allows up to 500 per avatar.
• Professional Voice Clone (PVC) — ElevenLabs' high-fidelity clone trained on 30 minutes to three hours of audio.
• Instant voice clone — a clone made from a few seconds to a few minutes of audio, quicker and less accurate.
• TTS — text to speech; software that reads text out loud.
• API key — a secret string that lets one service use your account on another.
• Credits — the currency HeyGen and others charge in; minutes of video cost different credits depending on the avatar type.
• Lip-sync — moving a face's mouth to match new audio.
• B-roll — cutaway footage laid over the main shot.
• MPS — Metal Performance Shaders, Apple's GPU backend that lets AI models run on Apple Silicon Macs.
• VRAM — memory on the graphics card; the thing local AI models run out of first.
• Watermark (audio) — an inaudible signal baked into generated audio so it can be identified as synthetic later.
• Affiliate link — a link that pays the person who shared it when you buy.

Sources

• How I Clone Myself With HeyGen & ElevenLabs, Julia McCoy (Dec 2025) — the tutorial, presented by the clone
• HeyGen pricing — current plans, credits, avatar slots
• HeyGen- Create your first Digital Twin — footage spec, steps, processing time
• HeyGen- Recording your consent video — consent video rules
• HeyGen- Digital Twin FAQ — avatar slots per plan
• HeyGen- Avatar Looks explained — adding outfits, 500-look cap
• HeyGen- Avatar IV complete guide — 31 and 16 credits per minute
• HeyGen- Integrate ElevenLabs and third-party voices — the import steps and API key permissions
• ElevenLabs pricing — Creator plan and Professional Voice Clone
• ElevenLabs- Professional Voice Cloning guide — audio length, format, verification, training time
• Synthesia pricing — Starter vs Creator, personal avatars
• Elai pricing — selfie avatar and voice clone add-ons
• Tavus pricing — custom replicas for conversational video
• Hume pricing — Creator plan and voice cloning
• Cartesia pricing — instant vs Pro Voice Cloning tiers
• Fish Audio plans — Plus plan and professional voice slots
• Chatterbox on GitHub — MIT local TTS with cloning and Perth watermark
• MuseTalk on GitHub — MIT real-time lip-sync
• InfiniteTalk on GitHub — Apache 2.0 audio-driven video
• LivePortrait on GitHub — video-driven portrait animation
• YouTube Help- Disclosing altered or synthetic content — when a label is required
• FFmpeg — the audio stitching tool
• DaVinci Resolve — free video editor