How to Narrate an Unreliable Narrator

An unreliable narrator is already working on the page. The reader gets one person's account of events, senses that something in it does not hold, and keeps listening anyway. When that book becomes audio, the common mistake is treating unreliability as something the voice has to add. It is in the sentences already, and the production job is to carry it without flattening it and without giving it away early.

The practical answer is to do less than you expect. Here is what that means in casting, in delivery, and in the parts of the setup that can quietly spoil a reveal.

Work out what kind of unreliability you have

A narrator who is lying to the reader is a different production problem from one who is honestly wrong, and both differ from one who reports events accurately while being completely wrong about what they mean. The signal in every case is a gap between what the narrator says and what the reader assembles.

A voice cannot perform a gap. It can only read one side of it. That is why a knowing delivery does so much damage: the moment the narrator sounds like someone with a secret, the listener stops assembling and starts waiting.

So the goal is a narrator who commits to their own account completely. The unreliability stays where the author put it, and the listener does the work they came to do.

Cast a voice the listener will believe

The instinct is to reach for something a little cold or a little off. It usually backfires. A voice the listener distrusts from the first paragraph has pre-announced the twist, and every later scene gets read through that announcement.

A warm, credible voice does more for this kind of book, because the listener has to be taken in before they can be surprised. Browse and preview options on the Voices page and pick the one you would trust with directions to the station. Our guide to choosing AI voices for characters covers how to audition a voice against a real passage rather than a demo line.

Test on a scene where the narrator is at their most persuasive. If the voice sells that scene, it will carry the rest. If you are still deciding how to frame the book, first-person versus third-person narration covers the casting difference in more detail.

The emotion control is scoped to dialogue, and that helps here

AudioProducer lets you attach an emotional tone to an individual dialogue line, so the same voice reads different lines with different inflection. It is set on the dialogue line itself in the editor.

That scope matters for this genre. An unreliable narrator's distortions are usually carried in narration rather than in quoted speech, which means the control most writers reach for to colour a suspicious sentence does not reach the sentences doing the work. For a book built on withheld information, that lands closer to a guardrail than a limitation, because per-line colouring on narration is exactly what telegraphs a reveal.

Where it earns its place is dialogue the narrator reports. If your narrator quotes another character and the quotation is one of the seams in the account, a tone on that line can hold the seam open. Put it on the other character's lines rather than the narrator's. Our post on adding emotion to an AI audiobook covers the control in general, and narrating internal monologue covers the related problem of separating a thought track from the narration around it.

Review both automatic passes more slowly than usual

Auto-Assign Sounds analyses the scene and places music, soundscapes, and one-shot effects to match what is happening. On most books that removes a large amount of production work. On this one it deserves a slower review.

The reason is clear once you see it. The feature reads the scene as the text describes it, and an unreliable narrator describes a scene that may not be the one that happened. Atmosphere matching the narrator's version can argue with the reveal later; atmosphere matching the real version gives the game away now. Keep the cues that support the narrator's account and pull anything that reads as commentary on it. A scene the narrator calls calm should probably sound calm, even where the reader is meant to suspect otherwise.

Auto-Assign Characters needs the same care. It tags every line by speaker in one click and leans on the punctuation conventions of the source, so it is least accurate on prose that drops quotation marks or attributes speech loosely. Literary and psychological fiction does that often. Read the tagged output and re-tag by hand where it slipped, because a line of dialogue folded into the narration track changes who the listener thinks is speaking. If the source drops quotation marks throughout, standardising the punctuation and re-running the pass is faster than correcting line by line.

The reveal chapter, and what should change in it

Less than you think. The reveal is written to reframe everything before it, and the reframing happens in the listener's head.

The one setting worth touching is pause length. AudioProducer gives you a project-wide default pause, a per-paragraph override, and inline pauses inside a paragraph. A slightly longer pause where the account collapses gives the listener the beat they would have taken on the page, where they would have stopped and looked back. Use the per-paragraph override so the rest of the book keeps its rhythm.

Resist re-casting the narrator for the final chapters. A voice change at the reveal tells the listener that the person they trusted was someone else all along, which is usually the opposite of what the ending is doing.

A lying character is a simpler job

Everything above assumes the narrator is the unreliable one. A reliable narrator with a lying character inside the story is a much simpler production, and separating the two before you start saves work.

There the narration track stays straight and the character gets their own voice like anyone else. The lie sits in the scene, other characters react to it, and the audio does not need to mark it. Books that alternate between a reliable narrator and an unreliable one are the case where a second voice earns its place, because the switch is structural. Separating two voices that sound too similar covers how to make that contrast land.

To set the whole thing up: import the chapter, run Auto-Assign Characters, fix the tagging, cast a narrator you believe, review the sound pass with the narrator's account in mind, then generate and listen to the persuasive scene rather than the reveal. Our guide to directing AI audiobook narration covers the review loop, and turning a psychological thriller into an audiobook covers the wider genre setup. You export an MP3 you own and publish it wherever you already publish. The free tier gives you 1,200 words with no card, which is enough to hear how your narrator sounds when they are being believed. Paid plans start from $39.99 per month.