Subtitles don't translate.
They follow.

How the live caption pipeline works

The pipeline has two stages. First, Automatic Speech Recognition (ASR) — a W3C-recognised speech-to-text mechanism — converts the speaker's audio into text in the source language. Second, Machine Translation (MT) renders that source-language text into each viewer's chosen target language. Both stages run during the call; the round-trip per clause is roughly one to three seconds. We deliberately wait for enough words to form a clean clause before showing the caption — accuracy is much better when the system isn't guessing mid-phrase.

If you've heard "AI-powered translation" or "neural translation engine," those are marketing labels for this same ASR-then-MT pipeline. We name the steps because honesty about the mechanism is the only way to be honest about the edges.

Per-viewer, not per-meeting

The speaker doesn't set a "meeting language." Each viewer picks their own caption language in their own client. In a Class 9 history class, the teacher speaks Hindi; one student reads English captions, another reads Tamil captions, the rest read the original Hindi as same-language captions (or turn captions off entirely). This is a deliberate design choice — a mixed-medium classroom is the default in India, not the exception.

Captions are client-side. They run in your browser. The speaker has no UI to set, no toggle to manage. The only person who needs to do anything is the viewer who wants a different language.

Captions and the Rights of Persons with Disabilities Act, 2016

Live same-language captions — English speaker rendered as English captions, Hindi speaker rendered as Hindi captions — are the live-equivalent of textbook accessibility commitments. The Rights of Persons with Disabilities Act, 2016, section 16, requires equal access to education for children with disabilities. UGC's 2022 Guidelines for Accessibility Standards in Higher Education Institutions ask HEIs to provide captioning for online lectures. LiveLoop's captioning is designed to fulfil this commitment in real time, not as a post-session add-on.

Visual handling: captions use a high-contrast overlay (configurable from the participant's caption panel), are labelled with the speaker's name so it's always clear who is talking, and can be enlarged on a per-user basis. We don't claim ADA or WCAG conformance because those are US/global frameworks Indian institutions aren't typically audited against — RPwD and UGC are.

Indian names, regional spellings, subject terms

The single most common complaint with English-default captioning tools is mis-rendering of Indian names. "Padmavati" comes out as "Padma watty." "Tamilakam" comes out as "tamil a calm." That's a fixable problem. The host can upload a custom vocabulary list before the session — student rolls, faculty names, institution acronyms (KVS, JNV, CBSE, ICSE), subject-specific terms (photolithography, syllogism, Tamilakam, Mauryan, Vishishtadvaita) — and the ASR stage recognises and spells them correctly from the first occurrence.

This is the same custom-vocabulary mechanism the transcription page uses for searchable post-session transcripts. The list is shared across both features; you upload it once, and both live captions and the saved transcript benefit.

What we do with audio. What we don't.

For captioning, audio is read into memory in short windows, the ASR stage produces text, and the audio buffer is discarded. The text — not the audio — is what gets translated and shown on screen. We do not write the audio to disk for captioning, do not retain it for "improving the model," and do not send it to a human reviewer.

If you do turn on cloud recording, that's a deliberate, separate switch with its own permission prompt and consent flow — recording stores an MP4. Captioning by itself, without recording, persists nothing.

Calendar permissions and OAuth scope (we only request calendar-event access, never inbox or contacts) are detailed on the security feature page. Captions don't touch the calendar surface.

Translation ≠ Transcription ≠ AI Assistant ≠ Insights

Four LiveLoop features touch language or intelligence in some way. They are deliberately not the same page:

• Translation (this page): the live, on-screen caption while the class is happening. Multilingual. Not archived (translation captions are not saved as a transcript).
• Transcription: the post-session searchable transcript archive. Click a sentence, jump the video to that moment. Source language, not multilingual rendering.
• AI Assistant: a post-session digest — key decisions, action items, named owners. Extractive (pulls sentences from the transcript) — not generative. Source-language summary, not multilingual.
• Insights: observable attendance and participation behaviour — who joined, who spoke, who voted in a poll. Not AI inference about engagement or mood. Deliberate boundary.

If you want raw caption-stream API access or webhook integration, that lives on the integrations page, not here. This page is the live, in-browser viewer experience.

Live captions vs. human interpreter vs. doing nothing

Use cases by institution type

K-12 schools (CBSE, ICSE, State Board)

Parent-teacher meetings where the teacher is most fluent in the regional language, and one or two parents have moved to a metro and are now stronger in English. Special-needs students who follow live class better with text running alongside audio. PTAs sharing announcements live with parents who travel out for work.

Colleges & universities

Multi-state cohorts in a single live lecture — a guest faculty from JNU lecturing in Hindi to a class with students from Tamil Nadu and Kerala on the call. Enabling Unit / Equal Opportunity Cell commitments under UGC accessible higher-ed guidelines.

Coaching centres (NEET / JEE / UPSC / SSC)

National-coverage coaching where the star faculty is bilingual but the student base spans Tier 2/3 cities. Captions let the same recorded lecture serve students who'd otherwise miss the regional-language nuances or technical terms in another language.

Skill-development & NCVET-affiliated training

Government-affiliated skill institutes pulling national-level trainers for sessions where the trainer's working language and the learners' reading language differ. The drop-off problem (silence in week one, withdrawal by week two) is the one captions most directly fix.