Turn an Audio-Only Reel Into a Talking Avatar: Models, Cost per Minute and the Rights Question (October 2026)
Some viral Reels are a voice over footage with no usable face: a podcast clip, a commentary, a voice memo. How to give that audio a speaker with OmniHuman 2.0, Kling Avatar or H3 Max Lip Sync, what each costs per minute, and why the audio is the part you must have rights to.
A large class of Reels cannot be remade by replacing a face, because there is no face to replace. The content is a voice: a podcast excerpt over waveforms, a commentary over b-roll, a voice memo with captions. If you want a version with yourself, or a character you built, as the speaker, the job is different. You take the audio and make a person deliver it. The models for that are audio-driven talking-avatar models, and the main question is not which one, but whether you have the right to the audio.
The job
- Get the audio. Your own recording, a voice you are licensed to use, or a text-to-speech track.
- Choose the speaker's image: a clear frontal photo of yourself, or a character you generated.
- Run an audio-driven model that animates the face (and in some cases the upper body) to the recording.
- Cut and caption. Join clips, add captions and the original visuals if you want them.
Which model, by length and budget
| Model | Price per second | 1 minute | Longest audio | Strength |
|---|---|---|---|---|
| MiniMax H3 Max Lip Sync | $0.05 (768P base) | $3.00 | 15 minutes | very long audio, lowest price |
| Kling AI Avatar v2 Standard | $0.0562 | $3.37 | about 5 minutes | volume, phone-size viewing |
| OmniHuman 2.0 | $0.105 | $6.30 | 5 minutes | body and gesture follow the delivery |
| Kling AI Avatar v2 Pro | $0.115 | $6.90 | about 5 minutes | a face that fills the frame |
Prices read on 18 October 2026. The H3 price is a base rate with unpublished multipliers above 768P. FlashTalk at $0.02 per second looks cheaper still, but it is fixed at a landscape 768 by 448 and cannot make a vertical video (why).
For a typical 30-second Reel, the choice is between OmniHuman 2.0 ($3.15) when the delivery matters and Kling Avatar Standard ($1.69) when it is volume. For a 10-minute talk, H3 Max Lip Sync ($30 at the base rate) is the only one that fits in a single call.
The picture matters more than the model
Evenly lit, frontal, the whole head in frame, a neutral expression, no sunglasses. In a vertical output the image should already be 9:16, because the video inherits the photo's shape. A restrained motion prompt such as "speaks calmly to camera, slight head movement" holds a face where "gestures energetically" does not. For anything over a minute, cut the audio into pieces of one to two minutes and join the results, because identity softens over time.
Making it more than a talking head
A talking head for 30 seconds is dull in a feed. Three ways to lift it without more generation:
- Cut away. Put the audio over the original b-roll or screen recording for most of the clip and cut to the avatar only for the lines that need a face.
- Captions in a bold style, which viewers read before they listen.
- Two crops: a tight shot and a wider shot of the same avatar, alternating at sentence breaks.
The rights question
The audio in a viral Reel belongs to someone: the speaker, the podcast, the platform's music library. Reusing it with a new face does not make it yours. Voice is also a likeness: cloning a specific person's voice, or putting a real person's recorded words onto a different face, is the use most likely to get a video removed and an account reported. The safe versions are your own voice and script, a licensed track, or a clearly fictional character with an original script. Our tools ask you to confirm the right to a source before they run, and you should ask yourself the same before you use the API directly.
What to expect honestly
Talking-avatar models are good at lip sync and calm delivery, and weaker at energetic gestures, fast speech, unusual accents and long clips. None of the models above is certified for identity, and the newer ones (OmniHuman 2.0, H3 Max Lip Sync) have not been through our own tests on real faces yet. Run ten seconds first.
Common questions
Can I use the Reel's original audio?
Only if you have the right to it. The audio belongs to the speaker, the podcast or the platform's music library.
Which model is best for a vertical result?
Any of the avatar models except FlashTalk, provided your photo is 9:16. The video follows the photo's shape.
How long can the audio be?
Up to 5 minutes for OmniHuman 2.0 and Kling Avatar, and up to 15 minutes for H3 Max Lip Sync.
How do I stop it looking static?
Cut away to b-roll for most of the clip, use bold captions, and alternate tight and wide crops of the avatar.
Can I clone a voice?
Only your own or one you are licensed to use. Cloning a specific real person is the use most likely to get a video removed.
How to do this without code: Giving a recording a speaker is the Animate photo tool with audio. In AutoHustle this is the Animate photo tool with a voice recording: audio in, a talking photo out (about 3 credits per second with the audio-driven model, quoted before you press Generate). See How to make a talking photo and Animate a photo.
Related reading: The AI Baby Podcast Video: How the Trend Is Actually Made · How to Animate a Photo Into a Talking Video With Lip Sync (2026)
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| MiniMax H3 Max Lip Sync | minimax/h3-max/lip-sync/image-to-video |
| OmniHuman 2.0 | bytedance/omnihuman/v2 |
| Kling AI Avatar v2 Pro | fal-ai/kling-video/ai-avatar/v2/pro |
| Kling AI Avatar v2 Standard | fal-ai/kling-video/ai-avatar/v2/standard |
| SoulX-FlashTalk | fal-ai/flashtalk |
Prices read on 18 October 2026. H3 Max Lip Sync: audio 5 s to 15 min. OmniHuman 2.0: audio 4–300 s.