AI Feature Issues

A verified email is required to generate summaries and use AI chat. Open the verification email Kalima sent when you signed up and follow the link, then reload the Studio. If you can't find the email, check your spam folder and resend it from your account. See Create Your Account for the full verification walkthrough.

This is expected. When you open a session that someone shared with you, the summary panel is read-only. You can read it and copy it, but Regenerate and dismiss controls belong to the session owner. You can still use AI chat on a shared session, and your chat history stays private to you. See Sharing a Session.

You've used up your plan's weekly allowance for that particular feature. Each of the three metered AI features (summaries, chat messages, and autocorrections) has its own separate quota, so running out of one doesn't block the others. Automatic session titles are free and never hit this limit.

To fix it, either wait for the weekly reset (quotas reset every Monday at 00:00 UTC) or upgrade to a plan with a higher allowance. Higher plans include substantially larger AI quotas, and Enterprise plans include unlimited summaries. See AI Usage & Limits and Plans & Pricing.

This is the short-term rate limit, not your weekly quota. All AI features, including read-aloud, share a brief per-user limit that smooths out bursts of requests. If you trigger many AI actions in a few seconds, you'll be asked to pause for a moment.

Just wait a few seconds and try again; the message tells you roughly how long to wait. Avoid rapid-fire clicking of Regenerate, repeated read-aloud, or sending chat messages in quick succession.

Yes. Every summary you generate, including each Regenerate, counts against your weekly summaries quota. Automatic session titles behave differently: they are written for free when a recording ends and never count against any quota.

The session has no transcript text yet. AI summaries are built from transcribed words, so you need to record a live session or upload a file and let it finish transcribing first. Once you can see transcript text in the Studio, generate the summary again. See Live Transcription or Transcribe an Uploaded File.

If a summary produces no new text for about 30 seconds, Kalima automatically stops it rather than leaving you on a frozen spinner. Just click Regenerate to try again. Summaries stream in as they're written, so on a normal generation you'll see text appearing within moments. A long silent spinner is the signal to retry.

Summaries save automatically and are stamped with the transcript they were generated from. If you edited the transcript (or cleaned it up with AutoCorrect+) after generating, the saved summary may no longer match. Click Regenerate so the summary reflects the current transcript.

Very long transcripts are summarized by sampling the beginning, a slice of the middle, and the end rather than processing every word, which keeps generation fast and reliable. The summary still captures the overall shape of the session. If a specific detail was missed, use AI Chat to ask about that exact moment, since chat works against the full transcript.

Summary depth depends on two things: how clearly things were said in the recording, and your plan. The summary stays faithful to the transcript and won't invent decisions, owners, dates, or quotes, so if a meeting was vague, the summary will be too. Higher plans with enhanced AI automatically use a stronger summary model for more detailed briefings, with no setting to toggle. Clean audio and session context also improve the transcript the summary is built from.

AI chat answers only questions about the current session's transcript. It politely declines unrelated requests, such as writing code, homework, essays, or fetching news or weather, and ignores attempts to override its instructions. Rephrase your question so it's about what was actually said in the session, for example "What did they decide about pricing?" or "Who owns the launch?".

Chat is grounded only in the transcript, so if the information wasn't said in the recording, it tells you it doesn't know rather than guessing. Confirm the topic was actually covered, and check the transcript directly. If the wording was misheard, fixing it with AutoCorrect+ can help chat find it.

Chat keeps the most recent part of a conversation in context, roughly the last 49 messages, and trims older turns to stay focused. Your saved history for the session is still there when you reopen it; it's only the working context that's capped. For a clean slate, use the clear / new-conversation action.

That's by design. Chat history is private per user. When you share a session, the recipient gets their own separate chat conversation on that transcript, so they can't see your questions and answers, and you can't see theirs. See Sharing a Session.

This is intentional. When you tick "also fix others," AI checks each other occurrence and only replaces the ones that are genuinely the same mistake. It deliberately skips places where the original word was actually correct in that context. If any remaining occurrences really are errors, fix them individually. See AutoCorrect+ (Mark & Fix).

If the marked text can no longer be uniquely located in its segment, usually because the transcript changed since you placed the mark, Kalima dismisses the mark rather than guess where to apply it. Re-mark the current text in the transcript and apply the correction again.

AI fixes in AutoCorrect+ are minimal by design: they correct what was misheard but won't add words that were never spoken (for example, they won't expand "Acme" to "Acme Corp"). If you need a different result, type the correction yourself instead of using the AI suggestion. The manual edit is a local change and gives you full control.

Typing a correction yourself is a free local edit. The metered AI calls are the "fix with AI" suggestion and the "apply everywhere" validation, which count against your AI autocorrections quota. To conserve quota, type fixes manually and reserve AI suggestions for tricky cases.

Read-aloud first tries your browser's built-in voices and falls back to Kalima's neural voices on the server for a wide set of languages. If a language isn't covered by the server fallback and your browser can't speak it either, read-aloud won't be available for that line. Try reading the original instead of the translation (or vice versa), since they may be in different, separately supported languages.

You can adjust playback speed anywhere from 0.5x (slower) to 2.0x (faster) in the read-aloud controls. See Read Aloud.

Only one segment plays at a time, so starting a new read-aloud stops the previous one. This is expected. You can also press stop at any moment to end playback. Note that read-aloud shares the same short-term AI rate limit, so rapid repeated requests may be briefly throttled.

All AI output (summaries, chat answers, automatic titles, and AutoCorrect+ explanations) is written in your account's display language, even when the transcript is in another language. If the language is wrong, change your interface language in your account settings and the next AI result will follow it. See Language Settings.