Improve Accuracy with Context
Give Kalima names, jargon, and a short description so it transcribes specialized content more accurately.
Session context is a short briefing you give Kalima before (or during) a recording. It is a sentence or two about the meeting plus the specialized words, names, and terms you expect to hear. It helps Kalima's transcription engine recognize vocabulary that generic speech models often get wrong, so your transcript needs less cleanup afterward.
Why context helps
Kalima already detects languages and adapts to most speech automatically. Where it can struggle is with words that sound ordinary but aren't: a colleague's surname, a product codename, a drug or diagnosis, a legal term, or an acronym specific to your team. Context tilts the engine toward the words you tell it to expect, without locking anything else out.
Think of context as telling a new note-taker, "We're discussing the Helios launch with Dr. Okafor and our partner Vantage Labs." Those few details are exactly what helps Kalima spell the names and terms correctly the first time.
What you can provide
When you open the session context dialog you can choose to Skip it, add a Description, or pick a saved Template, and in every case you can also add domain-specific terms.
A session description
A short free-text summary of the session: the topic, the setting, and anything unusual about the vocabulary. Keep it to a couple of sentences. The in-app quick description field holds up to 2,000 characters, which is plenty. Very long backgrounds give diminishing returns and aren't needed.
Key terms and names
List the words you most want spelled correctly: people's names, company and product names, acronyms, and technical jargon. These are the highest-value items to add, because they're exactly the words a general model is least likely to know.
If you're also translating the session, you can supply preferred translations for specific terms so they come out consistently in the target language rather than being translated literally.
Reusable context templates
If you record the same kind of session over and over, such as a weekly standup, a clinic, or a particular client's calls, save your context as a template and apply it with one click next time instead of retyping it. Saved templates are available on every plan, and there's no limit on how many you can keep.
Context guides recognition, it doesn't force it. Adding a name makes Kalima far more likely to get it right, but the engine still transcribes everything else it hears normally.
Add context to a session
You can add context as part of the setup wizard when you create a session, or open the context dialog directly from the Studio at any time.
Context applies to recovery too. If your connection drops mid-session, the stretch that gets re-transcribed afterward uses the same description, terms, and languages as your live session, so recovered text matches the rest of your transcript.
Store default context in a project
If a whole body of work shares the same vocabulary, set the context once at the project level instead of per session. Every new session created inside that project inherits the project's defaults, including its context, so each recording starts correctly configured.
A project can store either free-text context (up to 2,000 characters) or a saved context template; you pick one or the other. Editing a project's defaults requires the editor role or higher. This pairs naturally with a project's other shared defaults like language hints and audio quality mode.
Setting context on a project is the most reliable way to keep a recurring series consistent, because nobody has to remember to add the terms before each recording.
When context makes the biggest difference
Context pays off most when a session is full of specialized vocabulary. In medical and clinical settings it helps with drug names, diagnoses, and clinical abbreviations that generic models misspell. In legal and compliance work it captures case names, statutes, and specialized terminology. For product and engineering teams it handles internal codenames, feature names, and acronyms. And whenever named participants attend, it helps spell the attendees, clients, and presenters correctly.
Best practices
Get the most out of context:
- Add the handful of names and terms you care about most, rather than an exhaustive glossary.
- Spell each term exactly the way you want it to appear in the transcript.
- Keep the description short and specific. A couple of sentences beats a long essay.
- For recurring meetings, save a template (or set project defaults) so you never retype it.
- Combine context with expected-language hints for the strongest accuracy on multilingual or accented speech.
Limitations
- The in-app quick description field is capped at 2,000 characters, and concise context works best anyway.
- Within a project, free-text context and a saved template are mutually exclusive; choose one.
- Context improves the likelihood of correct recognition; it isn't a guarantee. For anything that still slips through, use AutoCorrect+ to mark and fix it afterward.
FAQ
No. Session context is available on every plan and doesn't change what your recording costs. It only helps the engine transcribe your specialized words more accurately.
Yes. Open the session context dialog from the Studio at any time to update the description and terms. Updated context applies going forward, including to any audio re-transcribed after a connection drop.
Expected-language hints tell Kalima which languages to listen for; context tells it which names, terms, and topics to expect within those languages. They work best together. See Languages & Detection.
The description is a short narrative of what the session is about; key terms are the specific words and names you want recognized. The description sets the scene, while the terms pinpoint the exact vocabulary that's easy to get wrong.
Yes. If a stretch of audio is re-transcribed after an interruption, it uses the same context, languages, and translation settings as your live session, so the recovered text stays consistent with the rest of the transcript.