Custom vocabulary and keyword boosting
In one sentence
Custom vocabulary tells a speech recognition system which unusual words to expect, such as your product names and industry terms, and keyword boosting raises the chance the recognizer picks those words when the audio is ambiguous.
Definition
Custom vocabulary tells a speech recognition system which unusual words to expect, such as your product names and industry terms.
Keyword boosting raises the chance the recognizer chooses those words when the audio is ambiguous.
This is the single highest-value, lowest-effort piece of configuration in a voice deployment, and it is routinely skipped.
The problem it solves
- General-purpose recognition models are trained on general language. Your brand name, product names, drug names, part numbers and local place names are underrepresented or missing entirely.
- Those are exactly the words a business voice agent most needs to get right, because they drive retrieval and action.
- A recognizer that meets an unfamiliar word substitutes the nearest familiar one, and does it with confidence.
How the mechanisms work
- Custom vocabulary supplies a list of terms the recognizer should be ready to output.
- Keyword boosting weights those terms so they win against phonetically similar alternatives while the audio is being decoded.
- Some systems take pronunciation hints alongside a term, which helps a lot for words whose spelling does not predict their sound.
- Boost strength is usually tunable, and over-boosting causes false positives, where the term turns up in transcripts of audio that never contained it.
Where it matters most
- Company and product names.
- Industry terms and abbreviations.
- Place names, local ones in particular.
- Alphanumeric identifiers, though these are often better handled by structured entry.
- Any term that shows up often in the knowledge base and rarely in general language.
Why the effect is out of proportion to the effort
- A word error on a filler word is invisible. A word error on the product name breaks retrieval, which produces a wrong answer, which is the failure the customer notices.
- Entity error rate predicts task success far better than overall word error rate, and custom vocabulary aims straight at entity errors.
Operational practice
- Pull candidate terms out of the knowledge base rather than compiling them by hand.
- Read the recognition logs for repeated substitutions, which reveal the terms you are missing.
- Revisit the list after any product or catalog change.
Common misconception
That a better recognition model removes the need for it. The best model available has still never seen your brand name. This is a lexicon problem, not a capability problem, and an afternoon fixes it.
Why it matters commercially
It is the cheapest quality improvement available in a voice deployment, and it lands on exactly the errors that cause visible failures, the misheard name or number that breaks the answer.
In voice specifically
A text chat receives the words exactly as typed, so there is no recognizer to bias and no proper-noun problem to fix. Speech introduces one, and custom vocabulary is the lever for it. The list you would never need for a typed form is the list a voice agent depends on.
Where AsqVox fits
Directly applicable. A voice agent doing true RAG over uploaded documents depends on recognizing the terms that appear in those documents, because those terms drive retrieval. Extracting custom vocabulary from the knowledge base is the natural way to do it.
Visual
The words the model has never heard
The best model available has still never heard your brand name.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Whisper Large-v3 records around 2.7 percent word error rate on clean benchmark audio and roughly 8 to 12 percent on real-world English.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - The clean figure is a public test-set result; the real-world half is a reported band rather than a single measurement, so quote it as a range and never as a provider figure.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate under production conditions.
5.26% to 6.84%vendor claimDeepgram Nova-3 reported figures, 2026 - Published by the provider. A production-conditions figure is more useful than a clean-benchmark one and still not comparable to another vendor number unless the test set and normalization match.
These are aggregate figures. Errors concentrate disproportionately on proper nouns, domain terms and numbers, which is why entity error rate predicts task success better than overall word error rate.
Errors cluster on entitiesindustry rangeReported speech-recognition error distribution, 2026 - Directional rather than a single measurement. It is the reason boosting a short list of business terms has an effect out of proportion to its size: it targets the errors that break the task.
No published benchmark quantifies the accuracy improvement from custom vocabulary across providers, despite it being standard practice and highly effective in reported experience.
-no reliable figureState the absence rather than borrowing a vendor uplift figure. The gain is real and reported everywhere; a comparable number is not.
No standard measure of entity error rate is published by providers, which is the metric that would demonstrate the benefit.
-no reliable figureWord error rate weights every word equally, so it hides the improvement custom vocabulary makes. The metric that would show it is not reported.
Examples
In practice
A retailer's voice agent hears its flagship product name three different wrong ways, so retrieval misses and the agent answers about some other product. Loading the product catalog as custom vocabulary with keyword boosting settles it: the term transcribes reliably and retrieval lands. Engineering time, one afternoon. Revenue effect, immediate.
The everyday version
Custom vocabulary is handing the system your words up front. Speech recognition learns from ordinary language, so it has never met your business name or your product names and will swap in something that sounds close. Give it the list and the guessing stops. It is the cheapest fix in the whole setup and the one most often skipped.
Usage
Who says it
- Speech engineers and voice deployment teams.
- The terminology varies: custom vocabulary, keyword boosting, phrase hints, biasing and custom language models all describe overlapping capabilities.
Where it turns up
- On a spec sheet it sits next to custom vocabulary support, boosting strength control, pronunciation hints, vocabulary size limits and update frequency.
- Support for these varies by provider, so verify the specific capability rather than accepting a yes.
Common misuse
- Skipping it on the assumption that a strong model does not need it.
- Over-boosting, which produces false positives where the term appears in audio that never contained it.
- Compiling the list by hand instead of extracting it from the knowledge base, then failing to update it when the catalog changes.
Questions people ask
What is custom vocabulary in speech recognition?
Custom vocabulary is a list of unusual words you give a speech recognition system in advance, such as your company name, product names, industry terms and local place names. General models are trained on general language and have never encountered those words, so without the list the recognizer substitutes the nearest familiar word. With the list it is prepared to output them.
What is keyword boosting?
Keyword boosting weights specific terms so they win against phonetically similar alternatives while the audio is being decoded. It works alongside custom vocabulary: the vocabulary tells the recognizer which words to expect and the boost decides close calls in their favor. Boost strength is usually tunable, and over-boosting causes false positives where the term appears in transcripts of audio that never contained it.
Does a better speech recognition model remove the need for custom vocabulary?
No. The best model available has still never seen your brand name, because it was trained on general language and your terms are underrepresented or absent. This is a lexicon problem rather than a capability problem, and it is fixable in an afternoon by supplying the terms. Errors concentrate on proper nouns and numbers, which are exactly the words custom vocabulary targets.
What should I add to a custom vocabulary?
Company and product names, industry terms and abbreviations, local place names, and any term that appears often in your knowledge base and rarely in general language. Alphanumeric identifiers are often better handled by structured entry than by boosting. The practical method is to extract candidate terms from the knowledge base rather than compiling them by hand, and to revisit the list after any catalog change.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.