This is Act II of the studio ladder: getting a character to speak on screen. We frame the whole craft as two roads and teach you to choose per shot.
Road A, native-audio models (one pass): prompt a generator that makes picture plus a synced dialogue track together. Google Veo 3.1 (synchronized dialogue, SFX, ambient, music at 48kHz, lip-sync reportedly under 120ms; Fast ~$0.15/sec, Standard/Quality ~$0.40/sec, Lite ~$0.03-0.05/sec no audio; Veo 3 shuts down June 30, 2026), ByteDance Seedance 2.0 (unified multimodal, leads audio-inclusive charts), Kling 2.6/3.0, Grok Imagine Video 1.5 (up to 15s, 480p/720p, multilingual lip-sync).
Road B, the decoupled pipeline: silent clip + separate voice + dedicated lip-sync pass on the mouth region. You own the exact voiceover, revise without re-rendering the picture, keep a clean stem.
How it works: phonemes map to visemes (b/p/m share one shape); neural models (Wav2Lip, 2020) skip visemes via a lip-sync expert discriminator. Audio-driven vs. performance-driven.
Voice: ElevenLabs - Free $0, Starter $5-6 (commercial rights + Instant Voice Cloning), Creator $22 (Professional Voice Cloning), Pro $99. Eleven v3 adds Audio Tags. Consent matters: Tennessee ELVIS Act (July 2024); NO FAKES Act pending; clone only with signed consent.
Lip-sync tools: Sync.so lipsync-2 ($0.04/sec) and lipsync-2-pro (diffusion, 4K, active-speaker detection); Hedra Character-3 (whole-face); Runway Act-Two (performance/gesture); Kling; HeyGen (dubbing at scale); D-ID (live agents); open-weights (Wav2Lip, LatentSync, LivePortrait).
Workflow: script -> voiceover first -> visual -> sync -> check on a big screen -> fix in the NLE. Cutaways hide a bad sync. Rough in cheap modes, finalize the keeper once.
News: Gemini Omni Flash public preview ($0.10/sec) and Nano Banana 2 Lite; Seedance 2.5 rollout window (30-second native, copyright cloud).