Talking-head video used to need a camera, a presenter and a quiet room. Two kinds of AI model now cover most of that work, and they are easy to confuse:
- Talking avatar (speech-to-video). You give it one photo and a voice recording. It creates the whole performance: mouth, expression, head movement, often gestures.
- Lip sync (video-to-video). You give it an existing video and a new audio track. It keeps your footage and changes only the mouth so the person appears to say the new words.
Choosing the wrong one is the most common reason people get a result they did not want. If you already have footage you like, you want lip sync. If you have only a picture, you want a talking avatar.
The models, and what each is for
TurboMax carries three models for this. All three take a voice track as an MP3 or WAV upload.
| Model | Starts from | Audio length | Output |
|---|---|---|---|
| OmniHuman 2.0 | One portrait photo | 4 s to 5 min | 720p |
| OmniHuman 1.5 | One portrait photo | Under 60 s | 720p or 1080p |
| Volcengine Lip Sync | An existing video (at least 360p) | Up to 240 s (Lite) / 210 s (Basic) | Your source resolution |
OmniHuman 2.0 is the one to reach for when the script is long: explainer videos, course modules, a product walkthrough read by a presenter. Five minutes of audio covers most of those in a single generation.
OmniHuman 1.5 is for short pieces where sharpness matters more than length, like a social clip, an ad hook or a greeting, because it can render at 1080p and keeps the framing of your photo.
Volcengine Lip Sync is for dubbing and corrections: translating a video into another language, replacing a fluffed line, or re-voicing a clip with a different read. It has two modes. Lite is meant for a single speaker facing the camera. Basic is meant for more complex scenes, and allows slightly shorter audio.
Both OmniHuman models also accept an optional short text prompt (up to 300 characters) to steer expression or gesture. Lip sync takes no prompt at all, since the audio is the instruction.
You can compare live per-second rates for all three on the models page.
Choosing a portrait that animates well
The photo decides most of the quality of a talking avatar. The model invents motion; it cannot invent information that is not in the picture.
Face the camera. A front-on or slightly angled face gives the model both eyes and the full mouth to work with. A strong profile leaves half the mouth hidden, and that half has to be guessed.
Make the face big enough. Head-and-shoulders framing works far better than a full-length shot. If the face is a small part of the frame, the mouth is only a few pixels wide and its motion will look smeared.
Keep the mouth visible and neutral. Hands, microphones, hair or sunglasses near the mouth confuse it. A closed or gently open mouth is the best starting point. A wide grin has to be "un-grinned" before it can speak.
Use a sharp, evenly lit photo. Hard shadows across the face and heavy compression artefacts get animated along with everything else.
Use a photo you have the right to use. Animating a real person's face to say words they never said is a serious matter. Use your own likeness, a presenter who has agreed to it, or a character you created. Impersonation and synthetic media passed off as real break TurboMax's acceptable use policy and, in many places, the law.
Preparing the audio
The audio drives every frame, so clean audio matters more than anything you can type.
Record voice only. Music or ambience under the voice makes the model try to lip-sync to the drums. Record dry, or use a vocal-isolated version of the track. Add the music back in your editor afterwards.
Trim the silence. Leading and trailing silence becomes seconds of a face doing nothing, and you pay for those seconds. Cut the file to start on the first word.
Speak at a natural pace. Very fast delivery, whispering and mumbling all give the model less to match. A clear, conversational read gives the most convincing mouth shapes.
Use a standard file. MP3 or uncompressed PCM WAV. TurboMax measures the length of the recording from the file itself. A file whose length cannot be read reliably is refused before anything is charged, rather than priced on a guess.
Mind the limits. Every upload on TurboMax can be up to 25 MB. The audio length limits are the ones in the table above.
How the cost works
All three models are priced per second of output video, and the output is exactly as long as your audio. That makes the cost easy to predict:
- Upload the recording. Its length is measured on the server.
- The studio shows the price for that length before you generate.
- Credits are reserved when the job starts and settled to what the provider actually charged when it finishes. If the job fails, the reservation is refunded.
The cheapest lever you have is the length of the audio. A 45-second script costs three-quarters of a 60-second one, so tightening the script saves more than any other setting.
For long-form work, split the script into sections, typically one per scene or slide. Generate each as its own clip. A problem in one section then costs you one short re-render, not the whole video.
A quick workflow
- Write and record the script, voice only, trimmed.
- Pick the source. A portrait means OmniHuman. Existing footage means Volcengine Lip Sync.
- Check the face. Front-facing, large in the frame, mouth unobstructed.
- Generate a short test. Ten seconds of the real audio shows you whether the face holds up before you pay for five minutes.
- Generate the full piece, then add music, captions and B-roll in your editor.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Mouth movement is smeared | Face too small in the frame | Crop to head and shoulders |
| Lips move out of time | Music or noise under the voice | Use a voice-only track |
| Face distorts on turns | Source photo in profile | Use a front-facing portrait |
| Long stretch with no speech | Silence left in the audio | Trim before uploading |
| Upload refused as unreadable | Unusual or compressed WAV variant | Export as MP3 or standard PCM WAV |
| Lip-sync result changes the scene | Wrong tool: avatar model used on a video | Use Volcengine Lip Sync for existing footage |
When not to use AI for this
A talking avatar is a good fit for explainers, training material, localisation and quick social content. It is a poor fit for anything where viewers will reasonably assume a real person was filmed saying those words, like testimonials, news or statements on someone's behalf. Label AI-generated presenters where your platform or your audience expects it.
Make a photo speak or re-voice a clip on TurboMax AI: OmniHuman 2.0, OmniHuman 1.5 and Volcengine Lip Sync, billed per second from one prepaid balance. No monthly subscription.