# MuseTalk

> Lip-sync an avatar video to speech audio - talking-head MP4.

Built-in lip-sync on your FlyMyAI account. Give it a reference avatar video
and a speech audio track, and it renders a talking-head MP4 where the avatar's
mouth follows the audio. This is a non-realtime batch step - it submits the job,
waits for it to finish, and saves the result as an `agent_file` with a
`public_url` you can feed into the next tool.

## What it can do

| Method | What it does |
| --- | --- |
| `lip_sync` | Lip-sync a reference avatar video (`video_url`) to speech audio (`audio_url`) and return a talking-head MP4 `agent_file` plus an `output_url` on object storage. Optional `bbox_shift` (mouth-region tuning, leave 0) and `fps` override. |

Both `video_url` and `audio_url` must be public HTTPS URLs reachable by the
service. The audio is typically the `agent_file.public_url` from a prior
`elevenlabs.text_to_speech` call.

## Typical pipeline

Lip-sync is usually the last step in an avatar pipeline:

1. `elevenlabs.text_to_speech` - turn your script into speech audio, then take
   `result.agent_file.public_url`.
2. `musetalk.lip_sync` - pass that audio as `audio_url` and your avatar base
   clip as `video_url`.

Call `lip_sync` on its own when you already have both URLs and only need the
lip-sync.

## How to get credentials

None - MuseTalk is a built-in tool. It runs on FlyMyAI infrastructure using
your account. Just enable it.

## Fields to fill in FlyMyAI

None.

## Troubleshooting

- **Input URL not fetched** - both `video_url` and `audio_url` must be direct
  public HTTPS links reachable by the service. Pre-upload via `download_link`
  or another agent tool if the source requires auth.
- **Job timed out** - long videos take longer to render. Keep the avatar clip
  and audio short, or split the audio into segments.
- **Wrong tool for realtime** - MuseTalk is a batch render. For a live realtime
  avatar endpoint, use `flymyai_deploy` instead.