Lesson 9 of 9 · about 24 minutes
Worked Example: Inside a 60-Second Podcast Intro
You'll be able to
- Read a 60-second intro timeline and see how a script is fit to it
- Follow how a synthetic narration and an instrumental bed are generated and proofed within a credit budget
- Explain how Volume Pins duck music under a voice, and work out the dB math
- Explain how Delivery Check confirms a podcast loudness target and what Apple's audio guidance asks of the files
The brief: an invented weekly civic podcast run by volunteers, the Harlan Valley Ledger, needs a 60-second intro. The hosts want a calm synthetic narrator so the intro sounds the same every week, an instrumental bed that does not sound like stock music, and a file that drops into any episode at the right loudness. The narrator must say it is synthetic. Here is how the job goes from blank page to upload, with every number worked.
Step 1: Plan the minute
| Time | Length | What happens | Music level |
|---|---|---|---|
| 0:00 to 0:05 | 5 s | Music open, establishing the theme | Full |
| 0:05 to 0:44 | 39 s | Narration, including the disclosure | Ducked 12 dB |
| 0:44 to 0:46 | 2 s | Music breath | Back up |
| 0:46 to 0:51 | 5 s | Tagline handing off to the episode | Ducked 12 dB |
| 0:51 to 1:00 | 9 s | Music out, fading over the last 3 s | Full, then fade |
Step 2: Fitting the script to the clock
Narration slot: 39 seconds. Allow about 2.5 seconds for the pauses at paragraph breaks, leaving 36.5 seconds of speech. At 150 words per minute, 2.5 words per second: 2.5 x 36.5 = 91 words. The tagline slot is 5 seconds, so 12 words at most.
Script (91 + 12 words)
Every week the Harlan Valley Ledger reads the agendas so you do not have to. Water rates. Road bonds. The library board's three-hour debate about parking. We go to the meetings, we request the records, and we tell you what changed and what it will cost you. No yelling. No spin. Just the documents, the votes and the people who cast them. I am your narrator, and yes, I am an AI voice reading a script written by our volunteers. The research is human. The voice is borrowed from a machine. [tagline] This is the Harlan Valley Ledger. Here is what happened this week.
Check against lesson 4: no numerals, no abbreviations, no names to mispronounce, sentences short enough to breathe. "Three-hour" is written out. The disclosure sits inside the narration where every listener hears it.
Step 3: Voice it
In Text2Speech, filter for narration voices with a calm, mid-low pitch and audition three previews. Generate the narration and the tagline as two separate files, so the tagline can be placed independently.
- Drafts: Turbo v2.5, stability 0.60, similarity 0.75, style 0.05, speed 1.00, pitch 0.
- Final: the same settings on Multilingual v2, exported as WAV.
Credit budget
The narration is 489 characters and the tagline 66, so 555 characters per full pass.
- Two drafts on Turbo at 0.5x: 2 x 555 x 0.5 = 555 characters.
- Final on Multilingual v2 at 1x: 555 characters.
- One fix, say regenerating the last paragraph (about 120 characters): 120.
Total: 555 + 555 + 120 = 1,230 characters for an intro that will play every week. That figure goes in the log.
Then the proof. The final narration file should run 37 to 40 seconds. If it runs 42, the fix is cutting a phrase from the script, not pushing speed past 1.05. After that, the file goes through Speech2Text for comparison with the script. On a 91-word script that takes two minutes and catches the dropped "and" your ears forgave.
Plan check: ElevenLabs' help center says free-plan content has no commercial license and must credit ElevenLabs in the title. A sponsored show needs a paid plan for the voice.
Step 4: Make the bed
Two routes, both instrumental.
- Voice Studio MusicGeneration: duration 1m, Instrumental only ticked. One generation, one track.
- MusicStudio: Instrumental (no vocals) mode, mureka-9.5, 2 variations (each billed). Use Extend or a bar cut if the length is off.
Prompt for either:
Instrumental Americana for a local civic podcast intro. Fingerpicked acoustic guitar, upright bass, light brushed snare, a little pedal steel. 96 BPM, steady and warm, clear melodic hook in the first five seconds, no big build, clean ending. No vocals, no synths.
"Clear melodic hook in the first five seconds" matches the 0:00 to 0:05 open in the plan. "Clean ending" asks for a rn eighth. Start at 12; go deeper if the voice is soft or the bed is busy, shallower if the music vanishes.
Step 6: Check loudness and fix it
In the Enhance tab, choose the Podcast target in Delivery Check. The help page lists it as -16 LUFS plus or minus 1, with a -1 dBTP true peak limit. That matches Apple's audio requirements, which ask for about -16 LKFS (the same measurement as LUFS) within 1 dB and true peak at or below -1 dBFS. Choose "As it would export" so the check measures the mixed tracks the way Export will, with applied pins, and click Check. It changes nothing.
Worked example: reading the report
First check: integrated loudness -19.2 LUFS, true peak -3.8 dBTP. Fail: 3.2 dB too quiet.
Plain gain would raise both numbers by 3.2: loudness to -16.0 (pass), but true peak to -3.8 + 3.2 = -0.6 dBTP, which breaks the -1 ceiling by 0.4 dB.
So export the mix first (Export with Mix all tracks ticked, WAV 24-bit) and open that file. Click the report's "Use this target in LUFS Normalization," which copies -16 into the LUFS Normalization slider without changing anything. Make sure Use limiter is ticked: the help page says that with it unticked, normalization stops short of the target if going further would push peaks past -1 dB, and with it ticked, it reaches the target while the limiter holds peaks at -1 dB. Normalize, then run Delivery Check again. A passing result looks like -16.1 LUFS and -1.0 dBTP.
Why the intro must match the episode: if the episode sits at -16 and the intro at -12, every listener reaches for the volume knob 60 seconds in, then again a minute later.
Step 7: Export
- Master: the normalized 24-bit WAV from Step 6, stereo, 44.1 or 48 kHz. Apple says it rejects mono WAV and FLAC for subscriber audio, so keep two channels even if the content is centered.
- Delivery: MP3 stereo at 128 kbps or higher; Apple recommends 128 to 256 kbps for stereo MP3. At 128 kbps a 60-second file is 16,000 bytes per second x 60 = 960,000 bytes, under 1 MB.
- Metadata: title "Harlan Valley Ledger intro," and a comment field reading "AI narration (ElevenLabs), AI music, see show notes."
Usually the intro gets mixed into each episode in your editor and the episode is exported once. Keep the WAV master for that, and never build next week's episode from last week's MP3.
Step 8: Disclose and log
The spoken disclosure is already in the narration. The full credit goes in the show notes once, in a template that repeats every week:
Intro narration: AI voice (ElevenLabs), script by the Ledger's volunteers. Intro music: AI-generated, edited by the Ledger.
Checklist before upload
- Script final saved, word count and character count recorded
- Voice name, voice ID, model and all tuning values recorded
- Music prompt, service, model, generation ID and date recorded; files downloaded before any 30-day link expiry
- Speech2Text comparison done; no mismatches
- Pins applied; music audible in the open and the outro
- Delivery Check passes Podcast: about -16 LUFS, at or below -1 dBTP
- Stereo 24-bit WAV master and MP3 delivery file exported
- Spoken disclosure present; show-notes credit added
- Plan for the voice and music services allows this use
Reporting angle
Podcasts and local radio increasingly open with a synthetic voice or generated music. Ask the producer whether the narrator is synthetic and where the show says so. Ask which service made the voice and the music, and request the terms under which the output may be published. Check the file against the host's published audio requirements, such as the loudness target the show claims to follow. If sponsors paid for the intro, ask for the order or contract, and ask whether sponsors were told the voice is not a person.
Research angle
Common mix-ups
- Claim: Louder intros grab attention. What the evidence shows: Apps normalize, and a jump in level annoys listeners. Match the episode's -16 LUFS.
- Claim: Lowering the whole music track is the same as ducking. What the evidence shows: SoundStudio's help notes that a flat cut leaves the music too quiet in the open and outro, where it carries the piece.
- Claim: A synthetic narrator needs no disclosure if a human wrote the script. What the evidence shows: Listeners assume a voice is a person. Saying otherwise costs one sentence.
