It started with a message. A friend who builds online courses wrote to me: “The course is all recorded, but I really don’t want to go back into a studio. Can you do something about it?”
He’d taken one of my classes and knew I’d been spending a lot of time with AI voiceover. This became my first full commercial project with it—a twelve-hour course split into forty-eight videos. This is my thirty-day project log: the quote, the failures, the back-and-forth. All of it, as it happened.
Days 1–3: Discovery, more fussy than I expected
By day one I’d learned that the hardest part of a voiceover project isn’t the voiceover. It’s figuring out what the client actually wants.
I gave him a list: Who’s the audience? Is it your own voice? Do you need to be able to change lines later? How long is each video? Do you have subtitle scripts already? His answers determined everything downstream. Using his own voice meant cloning it first. Needing to change lines later meant not locking everything into fixed audio files—keeping a re-generation workflow.
We settled on: a cloned version of his voice, forty-eight videos, about fifteen minutes each. Discovery took three days, and it was worth every hour. We had almost no rework later from mismatched expectations.
One question I didn’t think to ask but was glad I did: will these scripts change later? He said yes—the course gets revised twice a year. That answer pushed me to keep every segment as a separate file rather than one giant export, so future edits only regenerate the parts that actually changed.
Days 4–7: Auditions and dialing in the sound
Then the most important gate: three audition cuts for the client to choose from.
I deliberately made three versions—normal pace, slower, faster. “Does it sound right” is not something anyone can describe in words, so the ears have to decide. He picked the slower one, for a very practical reason: “Students take notes while listening. Any faster and they can’t keep up.”
That step taught me something important: the pace of a voiceover has to match what the listener is doing at that moment. Fast for commuters, slow for learners. That matters more than how good the voice sounds.
Once the audition was approved, I wrote every parameter—voice, speed, pitch, output format—onto a single “spec card.” Every video after that followed the card.
Before starting, I ran a stress test. I picked the hardest paragraph in the whole course—the one packed with proper nouns and numbers—and generated it first to confirm the settings could survive the worst case. If the hardest passage holds up, the rest is easy.
Days 8–15: Production, the most boring week
Production. Forty-eight videos, split into three or four chunks each—over a hundred segments.
That week was the same loop: paste the script, generate, listen, fix, save. Because the spec card locked down the variables, quality stayed steady, but it was tedious. I worked two fixed hours each morning and never pushed for more. With voice work, tired ears make bad calls, and you end up redoing everything.
The one failure that week: a video with a pile of proper nouns and English model numbers. I fed it in without tidying, and the model numbers came out spelled letter by letter. I stopped, made “put a space between English and Chinese” a standard step, and never hit it again.
To keep production moving, I made a progress board: forty-eight videos in a row, each one crossed off when it was done. It sounds silly, but long projects need visible progress. Facing forty-eight videos, it’s easy to feel like you’ll never finish—breaking it into “two today” shrinks the pressure and keeps you from missing any.
Days 16–20: Client feedback—editing was harder than voicing
The first delivery came back with three pages of notes.
Most weren’t mispronunciations. They were “that’s not quite how I’d explain it”—a concept he wanted framed differently, a line where he wanted an extra example. This is exactly where AI wins. With a human recording, those changes mean booking the studio again and rescheduling. With AI, I edit the script and regenerate that one segment.
I regenerated about thirty segments over five days, each in minutes. The client was startled by how fast changes could happen, and it was the first time I really felt that AI’s value isn’t “cheap”—it’s “flexible.”
The editing process changed how I worked, too. I used to ask “is this okay?” and get vague replies. Then I started offering two or three specific choices instead, like “should this line have an example, or be plainer?” The answers got much sharper. Turning open questions into multiple choice changes client communication a lot.
Days 21–25: Post-production, where the devil lives
For long content like a course, post-production is far more involved than single-clip voiceover.
Three things: normalize the loudness of every segment, even out the pauses between them, and add fades at the top and tail. Courses also have chapter openings, so I asked the client for a one-line chapter title and read it in the same voice as an audio divider.
One small thing I learned the hard way: file naming. With over a hundred segments, a messy naming scheme turns delivery into a disaster. I ended up using “course number - chapter - segment,” which the client could read at a glance.
I also backed up the masters. Every segment’s raw generation, post-produced file, and final delivery file went into three separate folders. With a project this size, the real fear isn’t doing it wrong—it’s finishing and not being able to find a file.
Days 26–30: Delivery and review
What I handed over: the complete audio, a mapping sheet showing which segment belongs to which video, and the spec card. I left the spec card with the client too, so if he wants to make small tweaks later, he knows exactly how I set things up.
Looking back after thirty days: the seven days spent on discovery and auditions were the best investment in the whole project. They front-loaded every possible round of rework.
Three things I’d change next time
Looking back after thirty days, three things stand out.
First, start discovery a day earlier. I began asking about requirements on day one, but I could have started during the proposal stage. Knowing one day sooner that he needed editable lines would have let me plan the workflow a day sooner.
Second, build one full-pipeline sample. I made three audition cuts, but the complete process—including post-production and delivery format—only got tested on a single video. Next time I’ll run one video end to end first, confirm post and delivery are clean, then start production.
Third, define revision rounds in the quote. I never specified how many rounds of changes were included, so when the client sent three pages of notes, I did all of them. Next time the quote will say “two rounds included, billable after that.” Not to be stingy—to protect both sides’ expectations.
The balance sheet: money, time, and one honest sentence
My fee for this project came to about a third of what a human voice actor would charge for the same volume. In time, I put in roughly thirty hours, including discovery and post-production. What the client got was a stable, endlessly editable, reusable audio asset.
The honest sentence: when you take on AI voiceover work, you’re not selling a voice. You’re selling a process. Clients want things on time, steady in quality, and changeable. Get the process right and the tool is just one part of it. If you want to understand where AI ends and humans begin, read this piece. To start with cloning, here’s the full record.
