Build a Pronunciation Glossary for AI Voiceovers
Create a maintained pronunciation register with exact spelling, locale, phonetic cue or approved audio, context, owner, and verification date for AI voice scripts.
A pronunciation glossary is a small, maintained table of words your AI voice gets wrong. You list the exact spelling, the locale, a phonetic cue or approved audio, the context, an owner, and a verification date. You check the glossary before every render. This method catches names, measurements, and borrowed words before they reach the final audio.
This is not a general audio quality checklist. It is a focused register for spoken output. It also differs from a full AI video with natural sound, which covers lip-sync and disclosure. And it is separate from what is voice cloning, which explains how a voice model is made. A glossary is the narrow layer that fixes how specific words sound.
Why a glossary beats fixing words one by one
A voice may read a name or abbreviation differently across scripts. For example, the word "read" changes sound with its tense. A glossary helps you record the intended sound in context. Without a glossary, you fix these errors by ear each time. That is slow and inconsistent.
A glossary gives you one source of truth. You record the correct spoken form once. Then you apply it to every script that uses the word. The W3C Speech Synthesis Markup Language (SSML) standard supports this idea. It provides a phoneme element for explicit pronunciation control and a sub element for spoken replacements. The standard notes that authors can replace written words with unambiguous spoken forms, such as writing "I will reed the book" instead of "I will read the book" (W3C SSML 1.1). A glossary is the human layer that feeds those controls.
What belongs in the glossary
Start with the words your voice actually gets wrong. Do not add every technical term. Add only the ones that fail a listen test. Common categories include:
- Names of people and places. Check the person's preferred spoken form.
- Borrowed words. Words from another language often follow the wrong rules.
- Measurements and units. "mg" versus "碌g" versus "Mg" sound very different.
- Acronyms and initialisms. "SQL" may read as "squeal" or "ess cue ell."
- Homographs. Words like "lead" or "read" depend on context.
For each entry, retain the exact approved spelling, including accents. Keep a separate field for how the term should sound. Do not alter the display name just to guide the voice.
The glossary worksheet
Use this table as your working format. All sample records, dates, audio filenames and approvals below are fictional. The sample pronunciation of Xylara is a team choice, not a language rule.
| Field | What to enter | Example |
|---|---|---|
| Written form | Exact spelling from the script | Xylara |
| Locale | Language and region tag, such as en-US or en-GB | en-US |
| Phonetic cue | A respelling or IPA string the voice understands | z-eye-LAR-ah |
| Approved audio | Link or filename of a correct recording, if available | audio/xylara-demo.wav |
| Context | Where the word appears, such as intro or product name | Fictional product launch |
| Owner | Person who confirmed the pronunciation | Fictional brand reviewer |
| Verified date | Last date the entry was checked | 2026-09-28 |
| Voice model | The specific voice or engine version tested | Voice A, v2.1 |
Keep the phonetic cue simple. A respelling like "z-eye-LAR-ah" is easy to read. IPA is more precise but harder to maintain. The SSML standard allows IPA values in the phoneme element, but support varies by processor (W3C SSML 1.1). Test which cue your tool accepts before standardizing on it. Do not assume one glossary entry fixes every voice. A cue that works for one voice model may fail for another.
A worked example
Imagine a fictional retail company called Harbor & Pine. They produce a weekly audio guide for store events. Their AI voice keeps mispronouncing three things: the product name "Xylara," the unit "sq ft," and the abbreviation "VIP."
The team builds a glossary with these entries.
Entry 1: Xylara
- Written form: Xylara
- Locale: en-US
- Phonetic cue: z-eye-LAR-ah
- Approved audio: audio/xylara-us.wav
- Context: Product name in event promos
- Owner: Brand manager
- Verified date: 2026-09-28
- Voice model: Voice A, v2.1
This is a fictional product name. The brand team chose the pronunciation "z-eye-LAR-ah" as a style decision. It is not a linguistic fact. The glossary records that choice so every episode matches.
Entry 2: sq ft
- Written form: sq ft
- Locale: en-US
- Phonetic cue: square feet
- Approved audio: none
- Context: Store size in event listings
- Owner: Copy editor
- Verified date: 2026-09-28
- Voice model: Voice A, v2.1
Entry 3: VIP
- Written form: VIP
- Locale: en-US
- Phonetic cue: vee eye pee
- Approved audio: audio/vip-demo.wav
- Context: Event access label
- Owner: Producer
- Verified date: 2026-09-28
- Voice model: Voice A, v2.1
Before each render, the producer checks the script against the glossary. For "sq ft," they replace the written form with "square feet" in the spoken script. The caption file keeps "sq ft" for readers. This split matters. The W3C captions guidance says captions are a text version of speech and non-speech audio needed to understand content (W3C WAI Captions). The spoken form and the readable form can differ. Do not paste markup blindly into a caption field.
For "Xylara," the team uses a respelling in the tool's pronunciation field. For "VIP," they use the approved audio clip as a reference and test the voice against it. If the voice still fails, they use a phonetic cue in the tool's supported format.
Edge cases to plan for
Names with multiple accepted pronunciations. Some names have regional variants. "Caroline" can be "CARE-oh-line" or "CARE-oh-linn." Ask the named person which they prefer. Do not infer accent from appearance. Record their answer in the glossary with the date. Get voluntary permission before storing an audio reference of a real person's name.
Words that change by context. "Record" is a noun or a verb. "Bass" is a fish or an instrument. Add a context column and create separate entries when the sound changes. The SSML standard notes that the same written word may have many spoken forms, such as "read" as "reed" or "red" (W3C SSML 1.1).
Fictional names in scripts. For an original fictional name, the project owner should approve the intended sound. Record that decision in the glossary so every episode matches.
Measurements with lookalike symbols. "碌g" and "mg" differ by one character. A voice may read both as "milligram." Test the expanded spoken form and verify the final audio. Keep the written form intact in captions.
Unsupported phonetic alphabets. Some tools accept IPA. Some accept only respellings. Some accept neither and need a sub-style replacement. This standard does not establish that Kyndrify supports SSML or any specific pronunciation field. Never paste markup into an unsupported field. Test one entry before building a large glossary. The SSML standard says support for specific phoneme alphabets varies and the default behavior when no alphabet is specified is processor-specific (W3C SSML 1.1).
How to maintain the glossary
A glossary rots if nobody owns it. Assign one owner per entry or per project. Set a review cadence. Monthly works for active projects. Quarterly works for slower ones.
When a new script arrives, scan it for glossary terms before rendering. When a render fails on a new word, add that word immediately. Do not wait for the next review. When a named person corrects their own name, update the entry and the verified date.
Keep the glossary next to the script library. A shared spreadsheet or a simple table in your project docs works. The format matters less than the habit of checking it.
Frequently asked questions
How many entries should a glossary have?
Start with five to ten words that fail often. Add more only when a render proves the need. A short, accurate glossary beats a long, stale one.
Can I use the same glossary for every voice?
Not always. Voices differ by locale and engine. A cue that works for an en-US voice may fail for an en-GB voice. Test each entry against the voice you actually use. Keep a locale column and a voice model column so you can filter.
What if the tool has no pronunciation field?
Use a spoken-script replacement. Write "square feet" instead of "sq ft" in the text the voice reads. Keep the original spelling in the caption file. This matches the SSML sub element concept, where an alias replaces the written text for speech (W3C SSML 1.1).
For consent and source-recording steps when using your own voice, see how to clone your voice. A glossary approval does not grant voice-cloning permission.
Start with one word
Pick the word your AI voice gets wrong most often. Fill in one glossary row. Test the fix in a short render. Listen on speakers and headphones. If the fix works, add the next word. Build the glossary one verified entry at a time. That is the whole method.
Disclaimer: This guide offers general steps for managing pronunciation in AI voice projects. Tool features and supported phonetic formats vary. Always check the terms and documentation of any service you use.
More from Kyndrify
Make your first video without filming.
Say what you need and the studio makes it: video, images, voices. Start free, no credit聽card.