Why do different characters in a localized ad need separate voice models?

Different characters need separate voice models because a voice is a per-character identity asset: sharing one model flattens personas, confuses viewers about who is speaking, and makes per-voice consistency checks meaningless. Generate and lock one voice per character, then clone each locked voice for that character's off-screen narration — with auto-regeneration if vocal drift exceeds 1.5%.

Treat each character's voice like you treat their face: a locked, reusable asset that belongs to that character alone. The invideo agent — an agentic video creation tool with the current generation models and voice tools built in — manages voices exactly this way in its localization workflow, and there are three reasons the separation matters.

A shared voice model collapses character identity. Two characters speaking through the same model sound like one person arguing with themselves — tone, pacing, and emotional register all converge. In localization this is compounded: a localized ad lands when the market hears someone native to it, not a translated track. As invideo's creative team puts it, "The thing that makes a localized ad land isn't translation. It's casting someone the market actually sees themselves in" — and that casting extends to the voice. One documented production maintained 3 distinct characters in a single ad, each with its own generated voice.

Consistency checks are run per voice, not per ad. The invideo agent's quality control compares audio between the original dialogue and each new B-roll narration segment with a 2-second audio-similarity comparison, and auto-regenerates any track whose vocal drift exceeds a 1.5% variance threshold. That check only works if each character maps to exactly one voice model — a shared model gives the drift comparison nothing stable to measure against.

Cloning depends on a clean per-character source. The workflow is: lip-sync one shot first to establish the character's voice model, then tell the invideo agent to clone and lock that voice for every shot where the character narrates off-screen. Before generating any dialogue shots, instruct the invideo agent to keep each voice consistent across that character's shots — it handles the rest, including generating the target-language voiceover (via ElevenLabs) with matched tone and lip-sync without an explicit prompt. If two characters drew from one model, the clone would blend both registers into every narration line.

The cost of doing this properly is small: across six localized ads in two markets (three ads, two markets each), one documented run spent ~$425 total — about $70 per ad and ~2.5 hours to recreate each — with separate locked voices for every character included in those numbers. Per-character voice modeling is also the wider industry standard: speaker-preserving dubbing APIs and multispeaker localization tools all assign one voice per speaker for the same reasons.

Watch some of these to see what works for you:

Also note that this voice will only be used for this particular character. We will generate different voices for each character separately.

— invideo's creative team