Last month I turned a 1,400-word article about home coffee brewing into a nine-minute audio episode. Start to finish, that job took 22 minutes, and roughly half of it was me listening back and catching my own mistakes. The tool doing the heavy lifting was PlayHT, and the workflow below is the one I’ve settled into after a dozen conversions like it.
The reason most people get terrible results from AI voice tools isn’t the voice. It’s the input. Paste raw blog text into a generator and you get something that sounds like a GPS unit reading a legal document out loud. Fix the script first and the same tool suddenly sounds like a person who read the piece and understood it. That’s the gap worth closing, and it’s the gap that makes the realism of modern AI voice generators genuinely unsettling the first time you hear it.
Step 1: Rewrite for the ear, not the eye
Spoken language and written language are different animals. Before anything touches PlayHT, run your article through a 10-minute edit pass:
- Break any sentence over about 18 words into two. Spoken sentences need to land.
- Kill parentheticals, footnote markers, and anything in brackets. Nobody can hear a bracket.
- Convert tables and bullet lists into full sentences. A voice reading bullet fragments sounds drunk.
- Read every hyperlink out loud as words or delete it entirely. “Visit example dot com slash pricing” is dead air.
- Spell out symbols: “percent” not %, “and” not &, “roughly 40 dollars” not $40.
My brewing article had a sentence like “Pre-infusion (3–5s) matters more than grind size — see §2.1 for the ratio table.” That became three clean sentences. Ugly on the page, perfect in headphones.
Step 2: Pick a voice by auditioning, not by browsing
PlayHT’s library is enormous, which is exactly why browsing is a trap. You’ll click 40 voices, fall in love with the most dramatic one, and then realise it sounds absurd narrating a recipe.
Do this instead. Take three sentences from the middle of your actual script, ideally ones with a question, a number, and a proper noun. Paste them into the same voice repeatedly and rate each one on two things: does it sound like it knows what it’s saying, and could I listen to this for 20 minutes without annoyance.
A few rules that have held up for me:
- Narration: pick mid-range, slightly understated voices. Warmth beats polish.
- Tutorials and explainers: conversational voices with mild imperfection win. Slightly-too-perfect is creepy.
- Anything with two speakers: pick voices that differ in pitch by a clear margin, not two similar-sounding narrators.
- Accents: match the audience, not your personal taste. A British voice reading US tax rules confuses people.
If you plan to keep this voice for months, clone your own rather than renting one. The process is less intimidating than it sounds, and cloning three voices in under an hour is a realistic benchmark for how quickly the setup goes.
Step 3: Fix pronunciation before you hit generate
This is where 80% of the quality complaints come from. Numbers, acronyms, brand names, and place names are the usual suspects.
Numbers
“1,200” might come out as “one thousand two hundred” or “one comma two hundred.” Write it however you want it said. For years, “2024” reads better as “twenty twenty-four” unless you’re listing a calendar date.
Acronyms
SEO, API, and NFT get spelled out. NASA, GIF, and SCUBA usually don’t. Test one instance of each acronym in your script before rendering the whole thing, then write the phonetic version that produced the right result. I keep a small find-and-replace list for every project.
Names and niche terms
Product names, surnames, and jargon are the worst offenders. Respell them phonetically: “SaaS” becomes “sass,” “Nguyen” becomes “nwin,” “gyro” becomes “yee-roh.” You’ll feel foolish typing it. You’ll feel worse hearing the wrong version in a published episode.
Step 4: Shape the pacing
Flat pacing is the second-biggest giveaway of AI audio. Fix it by not asking one generation to do everything.
Split your script into chunks of two to four paragraphs and generate them separately. You get natural reset points, and a bad chunk costs you 30 seconds to redo instead of the whole file. Between sections, insert a break tag so the listener gets a beat to absorb what they just heard.
Where you want emphasis, restructure the sentence rather than relying on emphasis tags. Short sentence first, then the point. “Here’s the part people miss. The grind matters more than the machine.” That rhythm reads as human because it is human.
Step 5: Listen at 1.25x before you publish
Speed playback exposes problems you’d otherwise skim past. Run through this checklist once, and only once, per episode:
- Any word that made you flinch? Add it to the pronunciation list.
- Any sentence that ran out of breath? Split it.
- Any section where your attention wandered? Cut 20% of it.
- Volume consistency between chunks? If one chunk is louder, re-render it.
- Opening line. Does it earn the next 30 seconds? If not, rewrite it and regenerate just that chunk.
That last point matters more than everything else combined. Audio has a brutal bounce rate in the first 20 seconds.
Step 6: Automate the repetitive parts
Once the workflow is stable, the boring parts become automatable: uploading finished audio, filing it by episode, writing show notes, pushing to your host. A trigger-and-action setup handles this well, and the pattern is the same one behind automating work that needs a bit of judgment rather than simple copy-paste jobs.
I keep the judgment calls manual: script edits, voice choice, the final listen. Everything downstream of a finished MP3 is automated.
When PlayHT is the wrong tool
It isn’t always the answer, and pretending otherwise wastes your time.
If you need full control over the model, want to run everything locally, and don’t mind a bit of setup, an open-source option is the better fit. OpenVoice runs on your own machine, which matters if privacy or cost-per-minute is the binding constraint.
If your real problem is verification rather than generation, you’re looking at a different category entirely. Detecting synthetic audio in the wild is a close cousin of cloning, built on the same technology, and no amount of voice tuning on your own project will help you there.
For most solo creators shipping regular audio versions of written content, though, PlayHT hits the sweet spot: fast enough to be practical, good enough that listeners stop noticing.
The pipeline, written down
Edit the script for spoken rhythm. Audition voices on real sentences. Neutralise numbers, acronyms, and names. Generate in two-to-four-paragraph chunks. Listen once at 1.25 speed. Fix and ship. Automate the filing.
Do it three times and you’ll stop thinking of it as an AI project. It becomes a production step, like adding featured images, that takes 20 minutes and doubles the places your writing can be found.