# How we use AI

AI runs through every stage of a live caption, from the first word heard to the last voice dubbed. Each step is a service you can switch on, tune or leave to a human.

## 1. Speech recognition

Speech-to-text models turn live audio into captions. Line 21 recommends the best engine for each language and lets you switch; some engines also tell speakers apart and follow conversations that change language.

- Engines: Speechmatics, Deepgram, Gladia, AWS Transcribe, Microsoft MAI-Transcribe
- Languages: 104
- Latency: 3s

[Speech recognition docs](https://line-21.com/docs/methods/asr)

## 2. AI Context

Give the AI the briefing a captioner would get: scripts, slides, agendas, and names. The AI Context Builder turns it into a description, names, and specialised language, which feed the proofreader and give speech engines a dictionary of words to expect.

- Engines: OpenAI, Anthropic

[AI Context docs](https://line-21.com/docs/ai-services/ai-context)

## 3. AI Proofreader

A language model checks every caption line against that context, fixing names, brands, punctuation, and filler words, and it can finish before the audience sees a word. It works on captions from people and machines alike.

- Engines: OpenAI, Anthropic
- Latency: <1s

[How the AI Proofreader works](https://line-21.com/features/ai-proofreader) · [AI Proofreader docs](https://line-21.com/docs/ai-services/ai-proofreader)

## 4. AI translation

Every caption line is translated live into as many languages as you need, with the engine chosen per language and custom dictionaries for brand names and terms.

- Engines: DeepL, Google, Azure, OpenAI
- Languages: 93

[AI translation docs](https://line-21.com/docs/ai-services/ai-translations)

## 5. AI dubbing

Speech-to-speech interpretation: the voice is carried straight into another language's audio track, without a round trip through text, for audiences who would rather listen than read.

- Engines: Gemini Live Translate
- Languages: 75
- Latency: 10s

[How AI dubbing works](https://line-21.com/features/ai-dubbing) · [AI dubbing docs](https://line-21.com/docs/ai-services/ai-dubbing)

## 6. AI call agent

A bot joins your Zoom call and transcribes it live, translating as it goes. Audio is used in the moment and never stored. Microsoft Teams and Google Meet are in development.

- Runs on the speech recognition engines above.

[AI call agent docs](https://line-21.com/docs/ai-services/ai-call-agent)

## People stay in control

AI is a set of tools in the workflow, not a replacement for the people who run it.

- **Human captioners, at no cost**: Bring your own captioners, respeakers, or Velotype writers; delivering their captions is free. ([Pricing](https://line-21.com/pricing))
- **Your choice of engine**: We recommend an engine per language, and you can test and switch it. ([Speech recognition docs](https://line-21.com/docs/methods/asr))
- **Checked before it airs**: Hold captions until the proofreader is done, so the audience only sees the final text. ([Before or after distribution](https://line-21.com/docs/ai-services/ai-proofreader#before-or-after))
- **Context you can edit live**: Change names and terms during the event; the proofreader uses them from the next line. ([AI Context docs](https://line-21.com/docs/ai-services/ai-context))

## Your data

Client content processed through Line 21 is not used to train any third-party AI model.

| Question | Answer |
| --- | --- |
| Used to train AI models? | No, for any vendor in the pipeline. |
| Exclude specific vendors? | Yes, per client engagement. |
| EU data residency? | Yes, depending on the speech and translation providers chosen. |
| Zero data retention? | Available for the language model providers. |
| Recordings kept? | 30 days by default, shorter on request. |

[Data handling and retention](https://line-21.com/docs/security/data-handling)