Listeners give you less time than you think. If the first 15 seconds don’t set up something worth staying for, the thumb moves. And if a sponsor read sounds like a contract being recited, listeners tune it out — and the sponsor doesn’t renew.
Those two slots happen to be the best fit for AI voiceover: a fixed script, repeated every episode, where a consistent voice is the point. Here’s how to write and produce both, and where you should still use a human.
What an intro actually has to do
A good intro isn’t an introduction to the show. It does three jobs at once:
| Job | Outcome | Common failure |
|---|---|---|
| Set expectations | Listener knows what the next 40 minutes delivers | Naming the show but not the value |
| Build recognition | Voice, music, and sentence pattern become a memory cue | New voice or music every episode |
| Reduce friction | New listeners know where to start | Demanding they hear 20 old episodes first |
The intro doesn’t need to be hyped. For interview and knowledge shows, a steady, clear, slightly curious delivery ages much better than radio-announcer energy.
15-second and 30-second structures
Most shows need two lengths: 15 seconds for the main feed, 30 seconds for trailers and cross-promotion.
| Seconds | Beat | What to write | Example shape |
|---|---|---|---|
| 0–3 | Hook | A specific problem or scene | “If you’re guessing which ad creative is working…” |
| 3–8 | Positioning | Who you are, who it’s for, what it solves | “This is Show Name — 30 minutes on one growth problem.” |
| 8–12 | Episode promise | What they get today | “Today: voiceover A/B testing, and one counterintuitive result.” |
| 12–15 | Action | Subscribe, or just start listening | “Let’s get into it.” |
| 15–22 (30s) | Proof | Audience size, guests, reviews | “Ten thousand marketers listen every week.” |
| 22–30 (30s) | Routing | Recommend a starter episode or newsletter | “Not sure where to start? Episode 12.” |
Two rules: the intro is for ears, not for a web page. Don’t spell the show name, don’t read slashes in a URL, and don’t stack three CTAs into 15 seconds.
Sponsor reads: smooth beats beautiful
The goal isn’t a pretty read. It’s that listeners don’t skip, and remember the code. Practice and most published guidance point the same way: host-read ads in the host’s own voice convert best. That’s exactly where AI fits — not replacing the host, but making the host’s delivery reproducible across every episode and every version.
A four-beat template
| Beat | Length | Content | Direction |
|---|---|---|---|
| Transition | 3–5s | Bridge from the previous segment | Same pace, no sudden announcer voice |
| Problem | 10–15s | What the listener is dealing with | Steady and specific, no exaggeration |
| Product + proof | 15–25s | How it helps, who uses it | One core feature only |
| Action | 8–12s | Code, URL, deadline | Slowest, repeat the key detail |
A 60-second read holds roughly 140–150 English words. Past that, listeners keep the last sentence and little else.
Five ways to keep it from sounding like an ad
- Keep the host’s sentence patterns. If the host says “I use this myself,” the AI version says it too.
- Leave one honest limitation. “It won’t write the strategy for you, but it saves a team producing three videos a week.” Specificity reads as sincerity.
- Vary the opening each episode. Same template, different entry point, so listeners don’t build an “ad alarm.”
- Say the code twice, at least five seconds apart.
- Duck the music 6–8 dB. If listeners have to work to hear the voice, they stop working.
Technical settings that matter
| Parameter | Suggested value | Why |
|---|---|---|
| Pace | 150–165 wpm English; 240–260 chars/min Chinese | Slightly faster than teaching segments |
| Sentence pause | 0.3–0.5s | Longer sounds hesitant |
| Paragraph pause | 0.8–1.0s | Lets attention reset |
| Integrated loudness | around -16 LUFS | Apple’s guidance is about -16 LUFS; Spotify normalizes near -14 LUFS |
| True peak | below -1 dBTP | Avoid codec distortion |
| Music under voice | -18 to -22 dB | The first two seconds can ride higher |
Platforms normalize loudness for you, so the number matters less than consistency: intro, read, and episode body should sit within 2–3 LU of each other. Big jumps make people reach for the volume slider.
Dividing the work between host and AI
| Approach | Who does what | Best for | Watch out |
|---|---|---|---|
| Host does everything | Host records intros and reads | Shows with a solid recording routine | Time cost per episode |
| Host + clone | Host records a sample; AI generates lengths and variants | Multiple episodes a week | Written consent and disclosure |
| Two voices | Host opens, AI narrator handles ads and promos | Hosts who don’t want to read ads | The two voices need separation |
The middle option is the most common. Record one good sample, then generate 15-second, 30-second, and code-swapped versions in minutes. Build the voice with voice cloning, then batch the variants in text-to-speech.
Consent and law: cloning a real person’s voice requires written permission covering use, term, and platforms. In the US, the FTC requires clear disclosure of paid sponsorships; AI-voice disclosure rules vary by platform and jurisdiction. These aren’t technical problems — they’re the show’s credibility.
Version control for reads
- Name files
<show>-<ep>-intro-15s-v2,<show>-<ep>-sponsor-<brand>-v1. - Keep the script and generation settings with each version.
- Before a sponsorship ends, know which episodes need pulling or replacing.
- Audit the back catalog once a quarter so expired deals stop airing.
A/B tests worth running
One variable at a time:
| Variable | Test | Metric |
|---|---|---|
| Hook style | Question vs. statistic | 30-second retention |
| Pace | +10% vs. baseline | Completion rate |
| Read length | 45s vs. 60s | Skip rate, code usage |
| Voice | Host original vs. clone | Listener comments, conversion |
Sample size matters. Smaller shows should compare across four to six episodes rather than declaring a winner from one.
Pre-publish checklist
- [ ] The first 15 seconds say what this episode delivers
- [ ] Show name spoken once, not spelled out
- [ ] Sponsorship disclosed before the read begins
- [ ] Promo code said twice, five seconds apart
- [ ] Intro, read, and body within 3 LU
- [ ] Transcript or captions updated
- [ ] Expired sponsorships handled in the back catalog
FAQ
Will listeners notice the intro is AI? Fixed-structure 15–30 second intros are the hardest to spot. Improvisation, laughter, and emotional turns are where it shows — keep those with the host.
Do sponsors mind an AI-read ad? Most care about conversion and brand safety, not the recording method. But if the contract specifies a host-read ad, follow the contract and clarify before signing.
Do I regenerate every episode? The intro can stay fixed for a season; regenerate only the episode-promise section. Sponsor reads get regenerated per code and flight, with the template unchanged.
Do I need to record a sample? Only if you’re cloning your own voice — plan on 10–30 minutes of clean audio. For a stable narrator, a platform voice works without cloning.
What about music licensing? Intro and read music needs a commercial license that covers podcast distribution, including streaming platforms. That’s a separate purchase from the voice.
Turn the intro into an identity
The intro isn’t the opening line. It’s the storefront. Nail the 15 seconds, fix the voice, change one small part per episode, and within a few episodes listeners will hear the first bar and know it’s time to pay attention.
Generate your first version with Soundwaver text-to-speech, keep the host’s voice consistent with voice cloning, and see plans on the pricing page. For scripts that convert better, read how to write voiceover scripts that convert.


