Transcription Accuracy
Improve Kalima transcription accuracy by diagnosing poor audio, wrong-language detection, missing words, disconnect effects, and cleanup options.
If your transcript has misheard words, the wrong language, missing names, or jumbled speakers, accuracy almost always comes down to a few fixable things: the audio quality reaching the engine, the hints you give it, and how speech is layered in the room. This page walks through each cause and the quickest fix.
Most accuracy problems are solved before you press record, by choosing the right audio mode, adding expected-language hints, and giving Kalima a little context about names and jargon. Use the checklist below to diagnose what you are seeing.
Start here: a quick checklist
The single biggest accuracy lever is clean input. A decent microphone close to the speaker in a quiet room, with the right quality mode, will outperform any after-the-fact correction. Fix the audio first, then add hints and context.
Diagnose by symptom
Background noise, echo, and a distant microphone are the most common causes of garbled text. Kalima offers two capture quality modes:
- Enhanced mode turns on echo cancellation, noise suppression, and auto-gain. It is the default and helps a lot with a laptop mic or in loud or echoey spaces.
- Raw mode records unprocessed audio and is best with a good microphone in a quiet room.
If your transcript is garbled, check that the quality mode matches the room you are actually in. Enhanced is the safer choice for almost every room. You can compare both on a sample with the Test modes option in the audio settings. Also move closer to the microphone and keep a consistent distance. See Audio Sources & Quality for the full breakdown.
When the padlock next to your languages is open, or no language is picked, Kalima detects the language itself, and that can stumble on accents, short phrases, or languages that sound similar. The fix is to pick the languages that are spoken and close the padlock.
Search for and add each language that will be spoken, up to three, and check that the padlock is closed before you start. Recognition then sticks to those languages, and switches between them mid-conversation still work. Every new session starts with the padlock closed, so this usually only happens when it was opened for a session. See Languages & Detection.
The picker offers about 60 languages, and you can add more than one. Learn more in Languages & Detection.
Generic recognition often misses people's names, company and product names, medical or legal terms, and other jargon. Give Kalima that vocabulary ahead of time with session context.
Open the session context dialog (or the Context step in setup) and add a short description of the session plus the specific terms and names you expect. This boosts recognition of exactly those words. The quick-description field holds up to 2,000 characters, but keep it concise, very long backgrounds have diminishing returns. You can also save reusable context templates for recurring meeting types. Full guidance is in Improve Accuracy with Context.
Kalima separates speakers automatically (Speaker 1, Speaker 2, and so on), but overlapping or very short speech is the hardest case for any engine, so it may occasionally misattribute a line. To reduce this:
- Encourage one person to speak at a time, cross-talk is the main cause of mixed-up attribution.
- Make sure each speaker is clearly audible to the microphone.
Once recording stops you can rename speakers to real names, and the numbering stays consistent across every take in the session. See Speakers & Diarization.
If you sit far from the microphone or use a quiet lapel mic, the engine may not hear you clearly. Adjust the recording gain before you record: open the gain settings (the gear next to the audio source) and boost or cut the level. The range runs from -12 dB to +12 dB in 1 dB steps, default 0, and your setting is remembered for next time. Nudge it up for quiet mics and down for hot, distorted input.
This is expected and not an accuracy problem. Live transcription shows a fast preliminary result first, then refines it into the confirmed version a moment later. Confirmed text can trail the live audio by up to about 2 seconds while the engine finishes each phrase. This brief delay actually improves the final wording by avoiding choppy, over-segmented results.
Kalima shows how confident the engine is in each word using subtle opacity. Fainter words indicate lower confidence and are worth a second look. Treat this as a review hint, not a guarantee. A cluster of faint words usually points to noisy audio.
Kalima captures audio from the moment you press record and buffers it, so the brief window before the connection completes is minimized. If you still notice clipped opening words, pause for about half a second after pressing record before you start speaking.
Review an existing transcript
Use faint-word confidence cues to find passages worth checking. For multi-speaker recordings, run Check voices after recording and reassign individual passages where needed. AutoCorrect+ word editing is still in development and is not available yet.
When accuracy looks worse near a disconnect
If a section of the transcript is marked as recovered or arrived late, it was filled in by Kalima's gap recovery after a network interruption rather than transcribed live. That recovered text uses the same languages, translation, and context as your live session and is merged into the right place. If a gap shows as failed, use the retry option. See Connection & Recovery Issues for more.
Still seeing poor accuracy?
If you have a good microphone, the right quality mode, expected-language hints, and session context in place and accuracy is still poor, the problem is usually the source audio itself, heavy background noise, far-field microphones, or multiple people talking over each other. Re-record under cleaner conditions where you can, and contact support with your microphone model, browser or desktop app, and a short example if you would like a closer look.