Dec 2, 2024 2 min read AI

By the Numbers: A Guide to Better AI Number Pronunciation

AI Voice Numbers Bug: How to Fix Eleven Labs' Biggest Challenge

Ever tried getting an AI voice to read "1246"? What should be a simple task turns into a garbled mess that sounds like a calculator having a breakdown. While Eleven Labs leads the AI voice revolution for content creators, podcasters, and video producers, I've discovered its surprising weakness: basic numbers.

TL;DR: AI voices struggle with numerical digits. This guide shows you exactly how to format numbers for perfect AI voice pronunciation, saving you hours of frustration.

The AI Voice Number Problem

It started with a simple test. I input "1246" into Eleven Labs' text-to-speech system, expecting a crisp "one-two-four-six." Instead, I got an incomprehensible jumble that made me question whether AI had forgotten how to count.

Understanding Why Numbers Are Hard for AI

According to Eleven Labs and WellSaid Labs documentation, AI voices consistently struggle with numerical representations. While neither company provides a detailed technical explanation of why this occurs, our understanding of language models suggests a plausible theory: AI voice models, trained primarily on natural language texts where numbers are typically written out as words, may have less experience processing numerical digits directly.

Think of it like this: when these models see words like "twenty-five" or "one hundred," they're working with familiar patterns they've encountered thousands of times in their training. But when they see "25" or "100," they deal with a different kind of representation altogether.

Regardless of the exact technical cause, one thing is clear from the documentation and testing: writing numbers as words dramatically improve pronunciation accuracy.

Quick Fixes for Common Number Types

Currency Problems Solved

❌ "$1,200,000"
✅ "one point two million dollars"

Phone Number Solutions

❌ "(555) 123-4567"
✅ "five five five, one two three, four five six seven"

Year Format Fixes

❌ "2024"
✅ "twenty twenty-four"

The Complete Number Formatting Guide

Copy this formatting guide for perfect AI voice pronunciation every time:

Number Formatting Rules:
1. Always write numbers as words when possible
   - "twenty-five" instead of "25"
   - "one hundred" instead of "100"

2. Format large numbers in spoken patterns
   - "one point two million" instead of "1,200,000"
   - "three hundred thousand" instead of "300,000"

3. Add natural pauses in sequences
   - Phone: "five five five, one two three, four five six seven"
   - Reference numbers: "one two four six"

4. Write years as typically spoken
   - "twenty twenty-four" instead of "2024"
   - "nineteen ninety-nine" instead of "1999"

5. Remove symbols and write them as words
   - "fifty dollars" instead of "$50"
   - "twenty percent" instead of "20%"

A Script-Processing Tip

For those working with larger projects, consider creating a preprocessing step that automatically converts numbers to their written form. While you'll still want to review the output to ensure proper context (especially for things like phone numbers versus quantities), this can save significant time in script preparation.