Custom Vocabulary for AI Voice Agents: Pronunciation Guide

Custom Vocabulary for AI Voice Agents: How to Get Names, Addresses, and Jargon Pronounced Right

An AI voice agent that mispronounces "Thompson-Reilly Orthodontics" or mishears the street name "Ashworth Close" costs trust and, eventually, calls. The short answer: modern voice AI platforms solve this deliberately, not by luck, through pronunciation dictionaries (for speech output) and custom vocabulary lists (for speech recognition) — two separate but complementary mechanisms you configure per assistant, without retraining the underlying language model.

Anyone deploying an AI phone assistant for a dental practice, a car dealership, or a real estate agency runs into the same recurring problems: proper names, addresses, brand names, and industry jargon. This article explains how the fix actually works under the hood, how the leading providers approach it, and how to manage it cleanly and centrally in Famulor.

Two Problems That Get Confused: Understanding vs. Pronouncing

A voice agent has two speech components, each with its own failure mode:

• Speech-to-Text (STT): The agent has to correctly recognize unusual words the caller says — an uncommon surname, a ZIP code combination, a product SKU. If the model mishears the word, the wrong term ends up in the transcript, and everything downstream (booking, CRM entry, routing) works off bad data.
• Text-to-Speech (TTS): The agent has to correctly stress and pronounce words it speaks itself — a company name, a medication, a technical term. This isn't a comprehension problem; it's an acoustic-correctness problem.

Both problems hit the same word categories: proper nouns (people, places, companies), acronyms and abbreviations, and domain-specific jargon — medical, legal, technical. Fixing only one side of the equation still leaves you with broken-sounding calls.

How the Leading Voice Engines Handle Pronunciation Control

Famulor is a vendor-agnostic platform connected to several STT and TTS engines at once. Each takes a different approach to controlled pronunciation:

The common thread: none of these providers require a full retrain. All four approaches are configuration, not model development — which makes them a one-week fix in practice, not a quarter-long project.

A Mini IPA Primer

The International Phonetic Alphabet (IPA) sounds intimidating at first but is manageable in practice: every symbol maps to exactly one sound, regardless of how the word is spelled. For the surname "Nguyen," a phoneme rule might look like <<ˈŋʷɪn>> — the stress mark (ˈ) flags the accented syllable, and the rest spells out the actual sound rather than the English letters. For most use cases, learning five to ten symbols — the primary stress mark (ˈ), a handful of vowel symbols, and common consonant clusters — is enough to cover the words that come up repeatedly. Teams that don't want to learn IPA from scratch can lean on a pronunciation reference tool or transcription helper — the important part is listening to the result in a test call before it goes live, not just reading the rule on paper.

Alias rules are the simpler starting point and cover most cases: instead of a phonetic notation, you substitute the word with a spelling the default engine already pronounces correctly — "SKU" becomes "skew," "Ste." becomes "Suite." Only reach for IPA once an alias rule still sounds audibly wrong.

Common Setup Mistakes

Four patterns show up again and again. First, teams build the dictionary once at rollout and never touch it again — new products, locations, or staff with unusual names quietly fall through the cracks. Second, teams fix only the TTS side, because they hear mispronunciations directly while testing, while STT errors sit invisibly inside the transcript. Third, rules get created per assistant instead of maintained centrally, so the same company name ends up with three different, partially outdated corrections across three language versions. Fourth, the dictionary gets built entirely from quiet office test calls rather than real ones — background noise, accents, and weak mobile connections produce different failure patterns in production than in a calm test environment.

Setting Up Pronunciation Control: Step by Step

1. Collect problem words from real conversations

Don't guess which words cause trouble — pull it from the data. Post-call analysis and call transcripts reliably show which terms get repeatedly misrecognized or mispronounced. Gather this list over two to four weeks of live operation before building a dictionary, so you don't spend time "fixing" words that were never actually a problem.

2. Sort by STT vs. TTS problem

Classify each problem word: is it said by the caller and misheard (STT), or said by the agent and mispronounced (TTS)? Many terms — your own company name, for instance — need both fixes at once.

3. Build TTS rules: alias or phoneme

Simple cases only need an alias rule ("Dr." pronounced as "Doctor"). Proper nouns with no obvious standard pronunciation — an uncommon surname, a vehicle model — need a phoneme rule written in IPA. That takes a bit of ramp-up but pays off for any term that shows up in nearly every call.

4. Add STT vocabulary

Enter rare proper nouns, product names, and jargon as keywords or keyterms. Short single words work fine as a plain list; multi-word expressions (street names with house-number patterns, for example) are more reliable as phrases than as individual words.

5. Validate with real test calls

Listen to the test calls, not just the transcripts — stress errors often don't show up in written text at all. Test deliberately across different speakers, accents, and background noise before going live.

6. Maintain centrally instead of duplicating per assistant

Once multiple assistants — different locations or languages, for example — use the same terms (your company name, common product names), a single shared, reusable dictionary beats a copy per assistant. It cuts maintenance work and prevents a fix in one assistant from being forgotten in another.

The Famulor Approach: Centralized Pronunciation Control, No Vendor Lock-In

Because Famulor is a vendor-agnostic voice AI platform connected to multiple STT and TTS engines simultaneously, you're not locked into a single pronunciation syntax. The knowledge base holds your company-specific terms centrally — product names, addresses, points of contact — and the agent pulls from it in every conversation. The voice library lets you pick the right engine and voice for each language and use case instead of being tied to one TTS provider.

For prompt-style reference pronunciations — "pronounce 'famulor' as 'FA-mu-lor'" — you use the AI Prompt Editor, which makes changes instantly testable without code or API calls. Because Famulor supports 40+ languages, this matters most for businesses with multilingual customer bases or locations in more than one country: a pronunciation fix for the English assistant doesn't automatically carry over to the German or French one — every language needs its own maintained list.

Real-World Examples by Industry

Riverside Family Dental, a 12-person practice, hit a recurring problem in its first two weeks after rollout: the assistant didn't reliably recognize "periodontal cleaning" and mispronounced the name of the practice's referring orthodontist, a hyphenated surname. After adding a jargon list to the dental practice knowledge base and a phoneme correction for the name, callback requests for clarification during appointment booking dropped noticeably.

A three-location car dealership group ran into a similar pattern: vehicle model names like "Ioniq" or "Cayenne" came out mis-stressed on the default TTS voice, confusing callers in test conversations. A small phoneme dictionary for the ten most-mentioned model names fixed the issue across all three locations at once, since the dictionary was shared across assistants instead of duplicated. More on vertical-specific setups is in the automotive industry overview.

For a real estate agent focused on a specific neighborhood, the bigger issue wasn't pronunciation but recognition: street names with unusual spellings kept getting misheard in the transcript, leading to callers being matched to the wrong listing. A keyterm list with the twenty most common street names in the target area solved it without touching a single TTS setting. More examples are on the real estate page.

What Bad Pronunciation Actually Costs

Pronunciation issues look like a cosmetic detail at first glance. In practice, they produce three measurable effects: a higher abandonment rate, because callers assume the assistant isn't legitimate and hang up; bad data, because misheard names or addresses get written into the CRM or calendar as-is; and more manual cleanup work, because staff have to correct misrouted or mismatched calls after the fact. Each of these translates directly into minutes per call, and minutes per call translate into cost. A call where the assistant has to ask a name twice or a misheard address gets corrected manually often runs 20 to 40 seconds longer than a clean call — and across a few hundred calls a month, that adds up to real extra talk time, either as queue delay for other callers or as extra capacity you have to pay for.

The setup effort is small by comparison: building out a company's ten to thirty most important terms typically takes one to two hours — a one-time job with light upkeep afterward. To get a rough sense of what cleanly configured pronunciation is worth in your own call volume and staffing costs, the Famulor ROI calculator below gives a quick estimate based on call volume, average handle time, and current per-call staffing cost.

Conclusion

Proper names, addresses, and jargon are why so many voice AI rollouts feel rough in the first few weeks — and why they're also fixable with relatively little effort. The fix is rarely a new model; it's a well-maintained pronunciation dictionary on the TTS side and a keyword or keyterm list on the STT side, both managed centrally instead of reinvented per assistant. Famulor bundles exactly these pieces into one platform: a knowledge base, a multi-engine voice library, a prompt editor, and 40+ languages — without locking you into a single provider. If your assistant is still stumbling over your own company name, the next useful step is a short live demo where we walk through your actual terms — company name, products, addresses — and show exactly how the setup works.

FAQ

What is a pronunciation dictionary?

A pronunciation dictionary is a list of word-to-pronunciation mappings that tells a text-to-speech engine how specific words should sound — either through phonetic notation (IPA) or a simple word substitution.

What's the difference between keyword boosting and a pronunciation dictionary?

Keyword boosting works on the speech-to-text side and improves how reliably a caller's word gets recognized. A pronunciation dictionary works on the text-to-speech side and controls how the agent itself pronounces a word. You need both to handle a word correctly in both directions.

Do I need a separate pronunciation list for every language?

Yes. Pronunciation rules are language-specific, since phonetics and stress patterns differ between languages. A company name needs its own rule in a German or French voice configuration even if it already works correctly in English.

How many words should a pronunciation dictionary contain?

There's no fixed number — any word that's repeatedly mispronounced or misheard is worth adding. In practice, 10 to 50 entries cover the most common issues for most small and mid-sized businesses.

Can I set up pronunciation fixes without coding knowledge?

For simple word substitutions, yes — most platforms, including Famulor, offer no-code interfaces for this. Phoneme-accurate fixes using IPA notation benefit from a basic understanding of the phonetic alphabet, but it's learnable with a bit of practice.

How often should I update the pronunciation list?

Right after rollout, review it after two to four weeks of live operation based on post-call analysis. After that, a check whenever you add new products, locations, or staff with unusual names is usually enough.

What happens if a word isn't in the dictionary?

The engine falls back to its default pronunciation logic, which handles most common words well. Errors mainly show up for rare proper nouns, jargon, or foreign-language terms that have no entry.

Does mispronunciation also affect recognition later in the same call?

Indirectly, yes: when the assistant mispronounces a term, callers often repeat or correct it, which lengthens the call and makes additional recognition errors more likely.