AI live avatar with measurable quality
Understand the Qwen3-TTS and MuseTalk live path, quality gates, operator controls, and practical limits.
Last updated: 2026-09-05
LiveCanvas turns approved text into speech with Qwen3-TTS and renders the selected avatar with MuseTalk. The live path is deliberately separate from slower image and motion generation: APIMart produces reusable assets asynchronously and is never placed in the per-frame loop.
The live path
- The operator approves an avatar version, matching voice, motion pack, scene, and runbook.
- Preflight validates consent, quality evidence, policy, subscription, available minutes, GPU capacity, and output routing.
- A signed coordinator grants one exclusive Session lease to one 4090-class worker.
- The worker maintains a 60–90 second queue of approved speech, renders 720 × 1280 video at 25 fps, and publishes H.264 video with AAC audio.
- The output page exposes a short-lived read-only capture route for OBS or a live companion.
Opening Studio or the output page does not start live-minute billing. Billing follows the server's authoritative LIVE state and stops during a confirmed runtime or media fault.
Quality is a gate, not a badge
The source video, voice preview, 30-second lip-sync preview, and motion clips keep versioned measurements. A valid preview must contain audio and pass duration, frame-rate, face-detection, A/V timing, lip-sync, and loudness checks. Missing measurements fail closed instead of being shown as zero or “passed.”
The final decision remains visual and audible. If a mouth, hand, expression, or loop looks unnatural, reject that version and create a replacement without overwriting history.
Safe operation
- Draft preview is private and does not mean the destination is live.
- Start is distinct from Resume and requires a fresh policy and billing gate.
- Pause preserves an intentional hold without letting the Agent continue speaking.
- Safe Idle provides a controlled non-speaking frame while waiting or recovering.
- Stop ends output, settles usage, releases the lease, and revokes the output token.
Worker commands use signed requests with replay protection. The browser never receives GPU credentials, provider credentials, or a reusable worker address.
Practical limits
V1 uses one GPU for one concurrent live session until capacity evidence supports a different limit. A good source recording and reviewed motion pack materially affect perceived realism; higher generation speed cannot repair an obstructed mouth, blurred face, or repetitive gesture. The operator must monitor the session and retain the ability to stop it immediately.