> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vowen.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Dictionary & Vocabulary
> Teach Vowen your specialized terminology, and the ways it gets misheard.
## What the Dictionary does
A dictionary entry does two separate jobs.
1. **It biases recognition.** The word is sent to your transcription engine as a hint, so the engine is more likely to produce it in the first place.
2. **It corrects known mistakes.** Each entry can carry a list of *misspellings*, the ways the word actually comes back wrong. Those are rewritten to your canonical spelling after transcription, deterministically.
The first job depends on which engine you use. The second works on every engine, always.
Typical entries:
* Product names (Kubernetes, PostgreSQL, Figma)
* People's names, especially unusual spellings
* Industry jargon and acronyms
* Company-specific terminology
## Adding words
Click **Dictionary** in the left sidebar. The **Dictionary** tab is selected by default; **Threads** and **Expansions** are the other two tabs on the same page.
Click **Add Word** in the window toolbar. Once you have entries, there is a second **Add Word** button beside the search box, and typing a term into the search box and pressing Enter adds it directly.
In the Add Word dialog, click **Add frequent misspellings** and fill in one row per variant. Press **Add another** for more rows.
Once your dictionary grows, use the **search bar** at the top of the tab to filter the list. Press `/` or `Cmd+F` (`Ctrl+F` on Windows) to jump to it.
There is no limit on how many words you can add, on any plan.
## Misspelling aliases
Biasing is only a hint. Even with "Kubernetes" in your dictionary, an engine will sometimes still hand back "cuber netties". A misspelling is the deterministic half of the entry, an exact phrase that Vowen rewrites to the entry's canonical text after transcription.
In the Add Word dialog the section is labeled **Frequent misspellings**, with the hint *"Anything you add here is auto-corrected to "Kubernetes" in your transcriptions."* Each variant goes in its own row, with the placeholder **How it gets misheard**.
| Entry | Misspellings |
| - | - |
| Kubernetes | cuber netties, cube nets |
| GitHub | github, git hub |
| Swaathy | Swati, Swathi |
How misspellings behave:
* Matching is case-insensitive, and anchored to whole phrases, so a misspelling of `cal` will not rewrite the middle of "calendar".
* Case-only variants are deliberately allowed. Listing `github` under `GitHub` is one of the most useful things you can do, and Vowen keeps it rather than deduplicating it away.
* Because this is a rewrite applied after transcription rather than a hint, it works on every engine, including the local models that take no vocabulary at all.
Entries that carry misspellings show a small pencil badge in the list, with the full rule in its tooltip.
This is the same replacement machinery [Threads](/features/threads) use. A misspelling belongs to a word you already have in the Dictionary and always resolves to that word's spelling. A Thread is a standalone rule that can expand a short phrase into anything.
## Which engines use your vocabulary
Most cloud engines can steer recognition towards your terms directly, and use your Dictionary well. A few can only take them as a loose hint. Vowen sends your list to each one and stays inside that provider's limits.
| Engine | How well it uses your Dictionary |
| - | - |
| Deepgram | Well, on both Nova-3 and Nova-2 |
| AssemblyAI | Well, for files and live dictation |
| ElevenLabs | Well, for files and live dictation |
| Speechmatics | Well |
| xAI | Well |
| Mistral | Well |
| Soniox | Well |
| OpenAI | Well, on `gpt-transcribe` and during live dictation |
| Cartesia | Well, live dictation only |
| Groq and OpenAI-compatible servers | Weakly, as a hint the model may ignore |
| Gemini | Weakly, as a hint, file transcription only |
| Sarvam | Weakly, as a hint, live dictation only |
### The exceptions
These engines receive **no** vocabulary at all. Your dictionary still corrects their output through misspellings, but it cannot influence what they produce.
* **Local Whisper Tiny and Base.** See the warning below
* **Local Parakeet** and **local Nemotron**, on both macOS and Windows
* **Sarvam** when transcribing a file or a meeting (live dictation does receive your terms)
* **Cartesia** when transcribing a file (only live dictation receives your terms)
* **OpenAI's diarization model.** Turning on Identify Speakers switches OpenAI to `gpt-4o-transcribe-diarize`, which takes no vocabulary of any kind, so your whole dictionary is dropped for that transcription
* **Gemini Live.** Live dictation ignores your terms; Gemini file transcription applies them
The split is not simply cloud versus local. Check your specific engine.
**Local Whisper applies vocabulary only on Small and larger.** Small, Small (EN), Medium, Medium (EN) and the Large models all use your terms. **Tiny and Base receive nothing at all** and are skipped entirely rather than nudged weakly.
Base was the recommended default for a long time, so if you have never changed your model, every word you add is currently being ignored at recognition time. Two ways out: move to Small or larger in [Speech models](/transcription/models), or attach misspellings to your entries, which are applied afterwards and work on every model including Tiny.
## How terms are prioritized
Every provider has a ceiling, and a long dictionary will hit it. Vowen decides what survives truncation by how much each term matters to you, not alphabetically:
1. **How often you actually say the term.** Usage is counted per term as you dictate.
2. **How recently you said it**, then how recently you added it.
A term you added a year ago and use daily outranks one you added yesterday and have never spoken. Ranking is automatic.
## Bulk import from CSV
Click the **Import from CSV** button in the search row (the download icon) to add many words at once.
* One word or phrase per row, in the first column
* A `word` header row is optional and detected automatically
* Duplicates are detected and skipped, and the preview tells you how many before you commit
The dialog offers a **Download sample CSV** link that saves `vowen-dictionary-sample.csv` with the expected shape.
CSV import handles words only. Misspellings have to be added in the Add Word dialog.
## Auto-add corrections
Auto-add corrections is a Pro feature, available on macOS and Windows.
When Vowen pastes a transcription and you then fix a word by hand, it can notice and offer to remember the fix. Nothing is ever added silently. Every proposal raises a confirmation banner next to the recording indicator, and you decide.
### Turning it on
Turn on **Auto-add corrections to Dictionary** in **Settings > General**. The row carries a **Beta** badge and reads:
> When you correct transcribed text, suggest adding the correction to your Dictionary for improved future recognition. Requires at least one AI provider to be configured.
The row stays greyed out and the switch stays disabled until you have connected an AI provider on the [Configure your AI](/ai-features/ai-setup) page.
### What the banner asks
Title: **Add to Dictionary?**, followed by the word you typed. Offered after a **single** sighting. A dictionary entry is only a hint to the recognizer, so a wrong one costs very little.
Title: **Always correct this?**, showing the rule as `"Swati" → "Swaathy"`. Offered only after **two** sightings, because the result is a permanent find-and-replace applied to every future transcript, and one odd correction is as likely to be you restructuring a sentence as a recurring recognition error.
Press **Add** to accept. The banner clears itself after 15 seconds if you ignore it, which leaves the proposal pending so it can come back.
### The rules it follows
* **One sighting** promotes a new word; **two** are required for a misspelling.
* The two sightings must fall within a **30-day** window. Older evidence is treated as stale and does not carry forward, because a problem that genuinely recurs recurs within a month.
* At most **five** corrections are taken from any single dictation. More than that is a rewrite the diff mis-read, not a set of spelling fixes.
* **Dismissing a proposal is permanent.** Closing the banner with its X records a refusal for that exact pair and Vowen never asks about it again. Only adding the word yourself in the Dictionary reinstates it.
## Related: pass screen as context
A separate feature reads names visible in your current window and passes them to your AI provider so it can repair ones the recognizer misheard. It attacks the same problem as the Dictionary, but it is not part of the Dictionary.
Find it on the **Configure your AI** page, on the **Configuration** tab, nested under "Enhance transcription with AI" and labeled **Pass screen as context**. It is macOS-only, and it reads the accessibility tree rather than taking a screenshot, so it needs no Screen Recording permission and makes no extra model call. See [Configure your AI](/ai-features/ai-setup).
## Dictionary, Threads and Text Expander
| | Dictionary | [Threads](/features/threads) | [Text Expander](/features/text-expander) |
| - | - | - | - |
| Triggered by | Speaking the word | Speaking the trigger phrase | Typing `:shortcut` |
| What it does | Biases the engine, then rewrites known misspellings to the canonical word | Replaces a phrase with any text, plain or rich | Expands a typed shortcut into any text, plain or rich |
| Trigger casing | Case-insensitive | Case-insensitive | Case-**sensitive** |
| Reach | Every engine for misspellings; recognition depends on the engine | Every engine | Every app, no dictation needed |
| Use when | A term is yours and you want it spelled right everywhere | You want a short phrase to become longer text | You want the same, while typing |
Start with the word alone. If it still comes back wrong after a few dictations, open the entry and add the exact wrong spelling as a misspelling. That converts an unreliable hint into a guaranteed result.
## Tips
* Add the exact spelling and casing you want. That string is both the hint and the replacement target.
* Add the full form and the abbreviation as separate entries when you use both.
* Skip common English words. Only specialized terms are worth the space in a truncated list.
* Keep misspellings specific. A very short misspelling like `cal` is anchored to whole phrases and will not corrupt longer words, but it will still fire on the standalone word "cal" wherever it appears.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.