Most production timelines treat voiceover before video as an afterthought. The voiceover gets added as the final layer, something you drop in once everything else is locked.
That order feels logical. It’s, in practice, one of the most expensive mistakes a buyer can make.
This isn’t a minor workflow preference. It’s a structural problem that affects cost, turnaround time, and the quality of the final product. Understanding why requires a look at how professional voiceover actually works, and where the AI tools most buyers are now using to build their productions break down in ways that aren’t immediately obvious.
The Foundational Problem: Video Is Rigid, Voice Is Not
A finished video has a fixed duration. Every cut, every transition, every moment of silence has already been decided. When you bring a voiceover artist in at that stage, you’re asking them to fit a living, breathing performance into a container that was built without them.
Professional voiceover isn’t a recitation. A trained voice artist controls pacing, breath, emphasis, and pause to serve the meaning of the copy and the emotional arc of the piece. Those elements aren’t decorative. They’re what makes the difference between a read that converts and one that sounds like someone reading a script.
When the video is already locked, those tools disappear. The artist isn’t performing anymore. They’re matching. Matching is harder, slower, more technically demanding, and more expensive than performing.
What AI Script Tools Get Wrong About Timing
Here’s where the problem has gotten significantly worse in the last two years.
A growing number of buyers are using ChatGPT, Claude, Gemini, or similar large language models to write their voiceover scripts. That’s understandable. These tools produce clean, professional-sounding copy quickly. The problem is what they can’t do: accurately estimate how long that copy will take a professional voice artist to read aloud.
The standard rule for spoken voiceover is well established. General spoken word delivery runs at approximately 150 words per minute, with a hard practical ceiling of around 180 words per minute at a very fast pace. Slower, more deliberate reads like corporate narration, meditation, or documentary can fall well below 130 words per minute. LLMs know this rule. They’ll tell you they know it if you ask. And yet the scripts they produce almost never conform to it accurately.
The reason isn’t that the model is ignoring the rule. The reason is structural.
Why the Math Never Works Out
Large language models generate text by predicting the next most likely token, which is a word or word-fragment, based on everything that came before it. They don’t have any internal experience of time. When a model writes a 60-second script and estimates its duration, it’s performing arithmetic on a word count after the fact. The model knows the rule the same way it knows any other fact stored in its training data. But knowing a rule and applying it as a real-time constraint during text generation are two entirely different operations.
The model is optimizing for the quality of the writing: coherence, completeness, professional tone. It’s not optimizing for a specific spoken duration. Word count is a byproduct of the generation process, not something the model is actively managing line by line. By the time it finishes writing and checks the math, the script already exists. The model has no mechanism to sense whether 160 words will take a professional 60 seconds or 80 seconds to deliver, because it’s never heard anything. It processes language as text, not as sound.
There’s also a subtler problem that compounds the error. LLMs tend to write copy that’s dense. Well-constructed sentences. Complete thoughts. Layered qualifications. That kind of writing looks efficient on a page but it’s systematically slower to deliver aloud than the casual conversational speech the 150 WPM baseline was built to measure. A sentence with three clauses and a parenthetical aside reads faster silently than it delivers in a professional studio read, where each clause needs a breath and a landing point.
The practical result: a script a language model estimates at 60 seconds will frequently run 70 to 80 seconds in a professional read, sometimes longer depending on the delivery style the project requires. For a meditation piece or a deliberate corporate narration, the gap between estimated and actual runtime can be even larger. And when the video’s already finished, that gap becomes a problem with a price tag attached.
The Time-Sync Premium: Why Fixing This Costs More
When a buyer arrives with a locked video and a script that doesn’t fit it, the work required to resolve that is categorically different from a standard voiceover session.
Standard voiceover: the artist reads the script, delivers a clean and well-paced performance, and the editor fits the video to the audio.
Time-synchronized voiceover: the artist has to record to a specific duration, often to the second. This typically requires:
- Multiple takes at different pacing adjustments to find what fits
- Back-and-forth communication with the buyer to identify which sections can be compressed and which can’t
- Potential re-editing of the script itself to bring total duration into range
- In some cases, pickup recordings and patch sessions to match specific cuts in the video
All of that takes more time, more sessions, and more communication. Professional voice artists, when they agree to take on time-synched work at all, charge significantly more for it. Rates for time-synchronized voiceover routinely run 50 to 100 percent higher than standard reads, and that’s before accounting for the additional revision rounds that typically result from trying to fit a performance into a container it wasn’t designed for.
The irony is that buyers who make the video first to save time often spend more of it, and more money, than they would have if they’d simply started with the voiceover.
The Workflow That Actually Works: Voiceover Before Video
The correct production order is: script, voiceover, video.
Write the script. Hire the voice artist. Record the voiceover. Then edit the video to match the audio.
This isn’t a new idea. It’s how broadcast production has worked for decades, because broadcast producers learned early that building video around audio is far easier than the reverse. A video editor can trim a shot by two seconds, extend a hold, or adjust a cut to match a beat in the audio. Those are minor adjustments. Rebuilding an entire performance to match a finished video isn’t.
When you have the voiceover first, you have a performance: pacing, emphasis, and timing that was crafted to serve the script. The video becomes the visual complement to that performance. That’s the intended relationship between the two elements, and it shows in the final product.
If You’ve Already Made the Video
This is the practical reality: many buyers reading this already have a finished or nearly-finished video and a script that may not fit it.
Here’s what to do.
Don’t guess. Don’t ask an AI tool to estimate whether your script will fit your video. As described above, those estimates aren’t reliable for professional voiceover purposes, for reasons that are built into how the technology works.
Get a real timing estimate. Send the script to a professional voice artist before committing to anything. An experienced VO artist can read through your script, tell you roughly where it lands in terms of duration, and flag the sections most likely to cause problems. This takes a few minutes and costs nothing at the inquiry stage.
Be upfront about the video. If the video is locked and can’t be re-edited, say so before getting a quote. This changes the scope of the project, the timeline, and the rate. A professional can work within those constraints, but needs to know about them before quoting.
Look for re-edit flexibility if you can. If your video was built in an AI tool, a standard editor, or even a platform like Canva, there’s often more flexibility than buyers assume. A section that runs five seconds long can frequently be extended with a hold on a visual, a slower transition, or a brief title card. Small adjustments at the edit stage can sometimes eliminate the need for a full time-sync session entirely.
A Note on Scripts Written by AI
If you used an AI tool to write your voiceover script, it’s worth having a professional review it before recording, and not just for timing. Scripts optimized for reading on a page often contain sentence constructions that are difficult to deliver naturally out loud. Run-on clauses, back-loaded qualifiers, and dense technical language that reads fine silently can create stumbling points in a live read.
A good voice artist will flag these issues. A great one will sometimes suggest small rewrites that preserve your meaning while making the copy easier to perform, which makes the final product sound better.
The Short Version
Record the voiceover first. Edit the video second.
If you’re using an AI tool to write your script, don’t trust its runtime estimate for professional voiceover purposes. The 150 words-per-minute rule is real and well established. The problem is that language models apply it as an afterthought rather than as a generative constraint, and they consistently produce copy that’s denser and slower to deliver than their estimates account for.
If your video is already finished and the script doesn’t fit, time-synchronized voiceover is available. But it costs more, takes longer, and requires more back-and-forth than a standard session. The earlier you bring a professional into the process, the less that problem costs you.
Donny Baarns is a professional voice talent and sports broadcaster with 13,000+ projects delivered for brands including Budweiser, New Balance, Verizon, Puma, and WebMD. He records broadcast-quality audio with standard 24-hour turnaround from his professional home studio. For project inquiries, contact donny@donnyvoice.com or use the contact form below.