• 01

  Write for the ear, not the page

  Long subordinate clauses kill listener comprehension. Use contractions ("it's" not "it is"), keep sentences under 18 words, and read your draft aloud before generating. If you stumble on a phrase reading it, the model will too.

• 02

  Force pronunciations with respelling and punctuation

  There is no SSML input, but you can steer pronunciation with creative spelling — write "ny-OO-trient" instead of "nutrient" if it lands wrong, and the engine will follow. Commas insert micro-pauses, em dashes insert longer ones, and ellipses produce a clear beat. Test the chunk, fix the spelling, regenerate.

• 03

  Generate in 4,000-character chunks for long episodes

  A single generation handles 5,000 characters. For full-length episodes, split at natural section breaks at roughly 4,000 characters per chunk — that leaves headroom for the engine's breathing punctuation. Splice the WAVs in your DAW with a 100 ms crossfade.

• 04

  Master to -16 LUFS mono / -19 LUFS stereo

  -16 LUFS integrated for mono is the Spotify and Apple Podcasts target. Stereo runs -19 LUFS by convention. Use a loudness meter (Youlean, free) on the master bus, push true-peak ceiling to -1.0 dBTP. Voice-only podcasts almost always end up mono — flatten the export and shave file size.

• 05

  Lock voice + speed + script-style for episode-to-episode consistency

  The Kokoro model is deterministic for a given voice and text. As long as you keep the voice name (e.g. River), the speed slider, and your script-formatting conventions identical, the narrator sounds the same six months from now. Save those three settings in a session-level note.

• 06

  Use a different voice for ad reads

  Listeners zone out when sponsorship copy sounds like the host. Run the host narration as Adam or River, switch to Bella or Daniel for the ad, and add a 200 ms music sting on the transition. The voice change is the cue that says "this is a sponsor" — IAB recommends explicit audio differentiation for ad clarity.

If you need a cloned voice that sounds like you specifically, ElevenLabs is the right tool. If you need consistent narration from a roster of preset voices with no character cap pressure, the math here is straightforward.