When synthetic narration works on a podcast
Synthetic narration suits the parts of a podcast that are not the main conversation: show intros, sponsor reads, chapter recaps and trailer edits. These are short, scripted and heard once, which is exactly where consistency beats character.
It is a poor fit for the conversation itself. Listeners form a relationship with a host over dozens of episodes, and an obvious change in speaker between episodes breaks that. If narration appears in episode one, it should appear in every episode.
Ad reads are the strongest case. A sponsor read is scripted, has to stay consistent across forty episodes, and is often recorded by someone whose voice does not suit the show. That is a real problem synthetic speech solves rather than a compromise.
Choosing a voice for long-form listening
The test is whether you can listen for ten minutes without noticing it. Brightness, exaggerated inflection and a narrow pitch range all fatigue faster than a natural speaking voice, because listeners read synthetic speech as slightly effortful even when they cannot say why.
Register matters more than accent. A mid-range voice sits comfortably for a whole episode, while a very high or very low voice draws attention to itself. For documentary and interview-adjacent formats, a lower register reads as credible without needing character.
Match the voice to the existing show rather than to where the audience is. A voice that changes between episodes is far more noticeable than an accent nobody comments on.
Writing a script that sounds spoken
Written prose does not read well aloud. Long subordinate clauses need punctuation the voice will misplace, and a sentence that scans fine on screen can become unintelligible when spoken at conversational speed.
Write for the ear: short sentences, one idea each, and a deliberate full stop where a speaker would breathe. Numbers and abbreviations should be written the way you want them said, because the model reads symbols literally. Write "20 percent" rather than "20%", and expect "Dr." to pause as an abbreviation rather than expand.
Read the script aloud before generating it. Anything you stumble over, the voice will stumble over too, and this check costs a minute.
Joining multiple sections without seams
Any narration over 2,000 characters has to be generated in sections and joined, and the joins are where synthetic audio gives itself away. Two things hide them: consistent volume across sections, and a pause long enough to sound deliberate.
Generate every section with the same voice and the same speed settings. Changing either mid-episode produces a shift that is immediately noticeable, and re-generating one section with different settings to fix it makes it worse.
The limit is per job, not per script. A thirty-minute episode is roughly nine sections, which is manageable joining work and much less effort than re-recording a line that did not land.
Step by step
Split the script into sections under 2,000 characters
Each job accepts up to 2,000 characters, which is roughly three minutes of narration. Write the script in sections that break at a natural pause, a topic change or a paragraph break, so the audio can be stitched back together without an audible seam.
Pick a voice that is comfortable over a long listen
Bright, heavily accented voices tire after twenty minutes. For narration, a mid-register American or British English voice with an even cadence is the safer choice. A deeper voice reads as more authoritative for documentary and true-crime formats.
Set the speed slightly below 1.0x
The default of 1.0x suits an advertisement. Long-form listening is usually better at 0.9x to 0.95x, which matches a natural speaking pace and reduces the fatigue that makes synthetic narration obvious.
Increase the paragraph pause at section joins
When joining separate jobs, a longer pause at the boundary makes the seam disappear. The paragraph pause defaults to 500 ms; raise it for a topic change and leave it alone for a sentence continuation.
Export WAV for editing, MP3 for distribution
WAV is uncompressed, so it survives being trimmed and rejoined in an editor without generational loss. Export MP3 only once the final cut is settled, since every re-encode degrades the audio.
If you already have a finished recording and want written notes from it rather than narration for a new one, see the podcast transcription guide instead.