Choose Languages, Models, and Custom Vocabulary

This guide is for people who already use profiles and want better transcription accuracy without changing their whole workflow. It shows how to tune a profile’s spoken language, pick the right recommended model, and add vocabulary terms for names or jargon that generic speech recognition often misses.

1. Open the profile’s language and model settings

Open Settings, select the profile you want to tune, then open the Language & Model section for that profile.

Screenshot placeholder: profile language and model section in Settings showing the spoken language picker, Top recommendation card, and Show all models button

This section is profile-specific. Changing it affects the selected profile only, which is why it works best after you already decided whether the profile is for writing, meetings, support, or a specific language.

2. Pick the spoken language you will actually say out loud

Start with Spoken language for this profile. This is the main language hint TypingVoi uses when recommending compatible speech-to-text models.

Use this rule:

• Pick the language you will speak, not the language you plan to edit into later.
• Keep separate profiles if you regularly dictate in different languages.
• Use Auto only when your workflow truly changes languages often enough that a fixed language becomes annoying.

An explicit language is usually the better choice when one profile exists for one stable task, such as Spanish, Support, or Client Meetings.

3. Start with the top recommendation before comparing everything

Below the language picker, TypingVoi shows a Top recommendation. This is the safest first choice for that profile based on the current language and profile context.

In practice, the recommendation area helps you answer two questions quickly:

• Is there already a strong model installed for this profile?
• If not, which model should you download first?

If the recommended row shows Use Model, the model is already installed and ready to assign. If it shows Download, TypingVoi is recommending a compatible model you do not have locally yet.

4. Understand installed versus downloadable recommendations

The recommendation list can include both installed and downloadable models:

• Installed recommendations are ready now and are the fastest path to testing the profile.
• Downloadable recommendations are compatible choices TypingVoi thinks fit the profile, but you need to install them first.

This is useful when you are tuning a new profile but do not want to leave the page to open Model Library first. You can download a recommended model directly from the profile settings, then come back and use it as soon as the download finishes.

If you want a broader view of shared storage, filters, or the Used by relationship across profiles, use Download and Manage Models.

5. Use Show all models only when the recommendation set is not enough

By default, the page focuses on recommended compatible models instead of the full catalog. Click Show all models when you need to compare beyond the shorter recommendation list.

That is most useful when:

• you want a different speed and accuracy tradeoff
• you are testing two compatible models for the same language
• you suspect the top recommendation is good but not ideal for your accent, speaking pace, or recording style

Do not treat Show all models as the default starting point. Most people get better results faster by testing the top recommendation first, then expanding only if there is a clear reason.

6. Expect tradeoffs between speed, accuracy, and workflow fit

Each model row gives you practical hints such as speed and accuracy ratings. Use them as comparison signals, not guarantees.

A simple way to decide:

• Prefer the recommended default when you want the least setup friction.
• Prefer a faster compatible model when short replies and quick turnaround matter more than perfect wording.
• Prefer a stronger compatible model when you are dictating drafts, names, or longer content that is expensive to fix manually.

If a profile feels consistently slow, inaccurate with proper nouns, or overkill for the task, that is usually a sign to test a different compatible model in the same profile rather than change unrelated settings.

7. Add vocabulary only for terms the model keeps missing

If the profile’s current final model supports vocabulary boosting, the profile also shows a Vocabulary card.

Screenshot placeholder: vocabulary editor expanded with several saved terms, one alias example, and the Save and Clear buttons visible

Use vocabulary for terms that matter repeatedly, such as:

• people’s names
• company names
• product names
• internal acronyms
• technical jargon
• place names your work mentions often

Good vocabulary lists are small and intentional. Add the terms you truly need, then test again. Do not dump long reference lists into the editor just because they might appear someday.

8. Format vocabulary entries the way the editor expects

Each line represents one entry.

Use these formats:

• AcmeCloud
• Khanh
• Nguyen
• OpenAI
• AcmeCloud: Acme Cloud, Acme clowd

The text before the colon is the canonical term you want TypingVoi to prefer. Comma-separated items after the colon are alias spellings or likely mistaken forms. This is helpful when a name sounds one way but should be written another way.

Best practices:

• Keep one concept per line.
• Add aliases only when there is a recurring wrong form to correct.
• Save after editing so the profile stores the normalized list cleanly.
• Stay selective. Best results are usually with a short list, not a giant glossary.

9. Understand the supported-model limitation before relying on vocabulary

Vocabulary boosting is not universal across every transcription path.

The current profile UI makes this limitation explicit:

• vocabulary is optional and only appears for supported local batch models
• live preview ignores the vocabulary list
• Whisper ignores the vocabulary list

That means vocabulary mainly helps the profile’s final supported model output, not every realtime or alternate backend path. If you expect a custom term to appear during live preview and it does not, that does not necessarily mean the vocabulary list failed. Check the final transcript result instead.

10. Tune names and jargon with a profile mindset

Vocabulary works best when it matches one specific workflow. Put the terms inside the profile where they belong.

Examples:

• a Support profile can include ticket product names and common customer acronyms
• a Client Meetings profile can include participant names and organization names
• a language-specific profile can include local names that a generic multilingual model often misses

If many unrelated terms are piling up in one profile, that is usually a signal the profile is doing too many jobs. Split the workflow into separate profiles instead of building one giant vocabulary list.

11. What to do when nothing looks compatible

If the page shows no compatible models, the problem is usually one of these:

• no suitable model is installed yet
• the chosen language is too narrow for what is currently available
• you need to download a recommended model first

Start by checking the recommended rows on the page. If the list is still too limited, open Download and Manage Models and refresh the shared library.

Related guides

• Create Profiles for Different Apps and Workflows
• Download and Manage Models
• Use TypingVoi for Multilingual Writing
• Troubleshoot Common TypingVoi Problems

Best next step: test one profile with an explicit spoken language, the top recommended model, and a short vocabulary list of names or jargon you actually use every week.