An AI Voice Over Workflow That Survives Client Review

A voice over workflow is a chain of custody: script, voice, take, master, captions, licence. The tool you choose affects one link in that chain. The other links decide whether the project ships on time. This guide lays out the seven stages we would use on a client job, with the evidence each stage should produce.

The tools named below are the ones in our 2026 ranking; the workflow applies to any of them.

Stage one: lock the script

The script is the specification. Before generating audio, freeze the words, the pronunciation notes, and the target duration. A script that changes after the first take forces a new credit spend and a new caption timing pass.

Write the pronunciation notes in the same document. Product names, acronyms, numbers, and foreign terms go in a list with a phonetic respelling. This list becomes the reference for every later revision.

Stage two: cast the voice

Casting is a shortlist problem, not a browsing problem. Pick three candidates, generate the same difficult sentence with each, and choose on intelligibility and tone. Do not choose on the demo reel: a demo is curated, and your script is not.

Catalog size helps here, but only up to a point. Murf states 200+ voices, Speechify states 1,000+, and a dated third-party table puts ElevenLabs at 10,000+. A larger shelf does not shorten the shortlist; a good filter does. If the tool does not publish metadata you can filter on, you will audition by scrolling.

Stage three: generate short takes

Cut on sentence or paragraph boundaries. Short files are easier to re-cut, easier to re-generate, and easier to compare across candidates. Name each file with the script version and the take number, for example ad-v3-line04-take2.wav.

Generate at the highest specification the plan allows. ElevenLabs states PCM at 44.1 kHz with a 192 kbps ceiling on Pro; Altered Studio publishes 48 kHz; Typecast publishes 44.1 kHz on paid plans and 16 kHz on free. Starting low and up-converting later does not recover detail.

Stage four: edit for performance

Synthetic reads often need breath, pace, and pause adjustments. Do those in the edit, not by regenerating the whole script. A single regenerated line is cheap; a full re-read is not.

Keep a change log. When a client asks for “the second version, but warmer”, the log tells you which take that was and what changed. Without it, you are guessing at revision history.

Stage five: QC against the picture

Listen with the visuals, not alone. Check five things:

• sync at the start and end of each line;
• plosives and sibilance on the loudest words;
• consistent pronunciation across takes;
• music and effects not masking consonants;
• caption timing against the final cut.

This is also where a sample-rate mismatch shows up. If the editor’s timeline expects 48 kHz and the file arrives at 44.1 kHz, the conversion is easy but the question is avoidable.

Stage six: captions and licence

A caption file is a deliverable, not an afterthought. WellSaid lists SRT and VTT from Starter, which means the timing can start in the voice tool. If your tool does not export captions, schedule a transcription pass before the final review, not after.

Licence evidence belongs in the same folder as the audio. Keep the plan name, the purchase date, a copy of the plan page, and the terms URL. If a cloned voice is involved, add the speaker’s written consent. A client may never ask, but the one time they do, the folder answers in seconds.

Stage seven: hand off

The final folder should contain:

1. the approved script with pronunciation notes;
2. the master audio at the highest published specification;
3. an MP3 review copy;
4. the caption file, if the tool or a transcription pass produced one;
5. the licence evidence;
6. a short note naming the voice, the plan, and the generation date.

That is the whole handoff. It is boring, and boring is the point: a producer should be able to rebuild the session six months later without calling you.

What the tools actually publish

The gaps in that table are the workflow risks. A tool with a beautiful demo and no caption export adds a step. A tool with a huge catalog and no licence line adds a conversation.

Where the human still decides

Three decisions cannot be automated: whether the read is right for the brand, whether the pronunciation is acceptable, and whether the licence covers the intended use. A model can generate a thousand options and still miss the line that makes an ad work. The workflow above exists to give those three decisions a clean place to happen.

If you are still choosing a tool, start with the 2026 results table and the methodology behind it. If you are choosing between two, the comparison posts go deeper on the pairs that matter.