Voicebook

An AI voice can only perform what the script tells it. Give it a script written to be read and you get a flat, confident reading with the acronyms said wrong.

Voicebook is the agent I built to fix the script before the voice is generated. It takes an untidy working script and returns one directed for ElevenLabs v3: acronyms respelled the way they are spoken, the order rebuilt for the ear, pauses written in, and a delivery cue on every line. A team gets a natural voiceover for each video without recording a voice actor for every one.

The effort moves upstream. Better direction before the voice is generated, instead of fixing a flat voiceover after.

What it does to a script

• Phonetic respelling for acronyms, symbols and technical terms.
• Structural cleanup into a spoken order, without losing the meaning.
• Pacing through line breaks, punctuation and pause markers.
• Delivery cues for warmth, confidence or reassurance.
• Clean output that pastes straight into ElevenLabs.

How the workflow came together

It started with listening. Every take that came back wrong showed me something the script hadn’t told the voice, and each one turned into a rule:

• It said acronyms as words. So every acronym and technical term gets respelled the way it’s spoken.
• It read the page, not the room. So the script gets reordered for the ear before anything else.
• It never stopped for breath. So the pauses are written into the text.
• It used one mood for everything. So every line carries a delivery cue.

Once the rules held up across a few courses, I wrote them into an agent, so anyone on the team could get the same read without knowing the rules.

I stopped fixing voiceovers and started fixing scripts.

Scale

More than 50 videos so far, across DIP IT, DIP Nursing, the Executive Functioning video series and other courses.

It was then extended for Build Your Path, a Women in STEM VR game. That meant several characters with their own identities and accents (Vietnamese, Bulgarian, New Zealand, Filipino and Nigerian), with every line split into named audio files that map to the game’s dialogue nodes.