How to Choose an AI Voice for Audiobook Narration: A Data-Driven Method

Picking an AI voice for a 30-second clip is easy. Listen to a few previews, choose the one you like. Picking a voice for a ten-hour audiobook is a different problem, because the thing that fails is not how the voice sounds. It's whether the voice is still the same voice in hour three.

Long-form TTS is generated in chunks, and most models start each chunk from scratch (we wrote about why AI narrators change voice mid-chapter). A voice that lands in the same register on every cold start will hold together across a whole book. A voice that improvises a new register on every take will sound like a rotating cast of narrators, no matter how good any single take is.

We learned this the expensive way. Here is the cheap way.

Screen the previews first, for free

Voice libraries publish a preview clip for every voice. Before spending anything on generation, run those clips through basic pitch analysis: median pitch, pitch spread, and whether the register wanders within the clip itself.

We screened 188 voices this way in an afternoon at zero cost. The screen will not pick your narrator, but it surfaces the red flags reliably. A voice whose pitch swings a couple of semitones inside its own 30-second demo will swing at least that much between takes.

The cold-start probe

The direct test is almost embarrassingly simple: render the same sentence three times, as three independent requests, with real text from your actual book. Then measure how far apart the takes landed.

Take the median pitch of each take. The gap between the highest and lowest, divided by the voice's base pitch, is the number that predicts seams. Our thresholds, calibrated against full-chapter renders:

• Under 3%: the register is locked. Joins will be inaudible.
• 3 to 5%: workable. Production mitigations absorb it.
• Over 10%: you will hear the voice change at every seam, and no pipeline will save it.

For scale, 6% is one musical semitone. The worst voice we probed drifted 16% between takes. The best held 1.3%. Same model, same settings, same sentence. The voice is the variable.

A probe costs three short renders per voice. A full chapter you end up throwing away costs a hundred times that.

Then listen like an editor

The numbers can't finish the job. We concatenate the three probe takes into one file, so every join in it is a simulated seam, and listen. Is it one person reading? Is the read alive or robotic? Does the tone fit the book? Would you spend ten hours with this narrator?

Both filters veto, in opposite directions. The most stable voice we ever measured, at 1.9% drift, was wrong for the book: gentle where the material needed epic. And some of the prettiest voices we auditioned drifted 12%. You need stability and quality at the same time, and you can test both in under an hour.

Which voices hold up

One pattern has repeated in every probe we've run: voices cloned from professional studio narration are far more stable than character voices. Clean, consistent source recordings give the model a tight target, and it re-lands on that target take after take. The folksy grandpa and the noir detective invite the model to re-improvise the character each time, and it accepts the invitation.

If you're casting a long-form narrator: start with voices tagged for audiobooks and narration, probe before you commit, and treat anything over 5% drift as a character actor rather than a narrator.

---

Every narrator on Lex passes this test before reading a chapter. Hear the ones that made it on our audio showcase.