AI talking avatar: turn a photo and an audio track into a lip-synced video

It does one job — make a face move and lip-sync to an audio track — across seven different underlying models, so you can trade off quality, price, input type and duration limit. Every model produces a video where the mouth and, to varying degrees, facial expression track the audio you provide. This covers narrated presenter videos, greetings, announcements and short talking-head clips from a single photo.

It does not write a script for you — you supply the audio (or, on PixVerse LipSync only, typed text that VdoBloom converts to speech) — and it does not generate a full scene or background; the person and setting come from your source photo or video. It also is not the same tool as VdoBloom's Spokesperson tab (which builds and saves a reusable synthetic persona by gender, age, ethnicity, style and accent) or the Testimonial tab (a UGC-style testimonial video builder) — those are separate, self-contained tools that happen to sit near this one in the sidebar and may use similar underlying avatar technology, but they are not this generic photo-plus-audio tool.

It also does not accept arbitrarily long audio on every model — the two 60-second-capped models (Omnihuman 1.5, Volcengine) will not process a full 3-minute voiceover; you would need Kling Avatar or PixVerse LipSync for anything past 2 minutes.

Kling Avatar Standard and Kling Avatar Pro (Kling AI's Avatar 2.0) take a photo, an audio file and a required text prompt describing the intended action. Standard renders at a fixed 720p, Pro at a fixed 1080p — the resolution is locked to the model, not a setting you choose. Both accept up to 5 minutes of audio, the longest ceiling of any model here, and are priced per second from a 300-second anchor (954 credits for Standard, 1910 for Pro at the full 5 minutes).

Omnihuman 1.5 and InfiniTalk are photo-plus-audio talking-avatar models aimed at realistic lip sync at shorter lengths: Omnihuman caps at 60 seconds of audio, InfiniTalk at 120 seconds, and InfiniTalk's price depends on whether you generate at 720p or 480p. WAN Speech-to-Video Turbo is a third photo-plus-audio option with its own exposed controls (frames, guidance scale, safety checker) for more advanced tuning, also capped at 120 seconds and priced by resolution tier.

The remaining two models take video instead of a photo. Volcengine Video-to-Video Lip Sync re-syncs an EXISTING video's mouth movement to a new audio track — useful for redubbing footage you already have — capped at 60 seconds and offered in a Lite (faster) or Basic (higher quality) mode. PixVerse LipSync also starts from a source video, but is the only model in this group that can generate its own audio: instead of uploading a file, you can type up to 200 characters and pick a TTS voice, and the tool synthesizes the speech before lip-syncing it to the video.

Presenter and spokesperson-style videos from a single headshot: narrate over audio you recorded or had generated elsewhere, and the photo delivers it on camera. This is the most common path across all five photo-based models.

Redubbing an existing video into new audio without reshooting — record a corrected voiceover and re-sync it to the original footage with Volcengine, keeping the visuals but fixing or replacing what was said.

Quick talking clips with no audio recording step at all, using PixVerse LipSync's type-and-pick-a-voice path, when you don't have a microphone or a script recorded but do have a short line of text and a source video.

Long-form narrated content up to 5 minutes on a single avatar image, which only Kling Avatar Standard or Pro support among these seven models — the others top out at 1-2 minutes.

Every model here is billed per second of audio (or, for the video-based models, per second of the target audio track), so the same 30-second clip costs a different amount depending on which model renders it. The table below gives representative costs at three audio lengths across all seven models, using each model's own duration cap and rate.

Trim your audio to the length you actually need before uploading — since every model bills per second, a 90-second file with 60 seconds of silence still costs for the full 90 seconds.

Use a front-facing, well-lit, single-face photo for the five photo-based models; the lip-sync and expression quality depends on how clearly the mouth and jawline are visible in the source image.

For the two Kling models, write a specific action prompt ("the man is speaking directly to the camera, calm expression") rather than leaving it generic — the prompt is required and directly shapes the output.

If you need more than 60 seconds of audio, skip Omnihuman 1.5 and Volcengine and use Kling Avatar, InfiniTalk, WAN Speech-to-Video Turbo or PixVerse LipSync instead, since those two are hard-capped at 60 seconds.

For a quick, low-cost test before committing to a long audio track, PixVerse LipSync's per-second rate and 5-credit minimum make it the cheapest way to preview lip-sync quality on a short clip.