Clarity AI Transparency
In short: every AI feature in Clarity runs on your iPhone. Your audio and words are never sent to a server, never used to train AI models, and Clarity does not identify who anyone is.
1. Where Clarity uses AI
- Speech recognition turns speech into punctuated text. Model: Parakeet Unified 0.6B (English), licensed by NVIDIA Corporation under the NVIDIA Open Model License, in a Core ML conversion by FluidInference, running on your iPhone through the open-source FluidAudio library.
- Speaker detection tells different voices apart so lines can be labeled “Speaker 1”, “Speaker 2”, and so on. Model: Streaming Sortformer v2.1, licensed by NVIDIA Corporation under the NVIDIA Open Model License (Core ML conversion by FluidInference), on your iPhone.
- Summaries (optional, only when you ask) write short key points for a saved conversation. Model: Qwen3.5 0.8B by Alibaba Cloud, on your iPhone through llama.cpp.
Clarity’s live-caption pipeline is adapted from the open-source EchoAssist project. Full credits and licenses: Open-Source Licenses.
The models are downloaded once from our servers, then run entirely on your device. Model versions are fixed for each Clarity release; if we change a model, we will note it in the release notes.
2. What Clarity’s AI does not do
- It does not send audio, transcripts or summaries to us or to any AI provider.
- It does not use your conversations to train or improve AI models.
- It does not recognize or identify people. Speaker labels only separate voices within a single session; no voiceprints or biometric profiles are created or stored, and labels start fresh every conversation. Names appear only when you type or choose them.
- It does not record audio. Audio is processed in memory and discarded.
3. Limitations
AI can make mistakes. Accuracy is lower with background noise, music, distance from the phone, people talking over each other, strong accents, uncommon names and technical terms, or quiet speech.
- Captions may miss or mishear words, especially names, numbers, dates and medicines.
- Speech may be attributed to the wrong speaker, particularly when voices are similar or people speak at the same time. Clarity can tell apart up to four voices at once.
- Summaries are produced by a small on-device model. They can leave out important points, attribute something to the wrong person, or state details that were not said. Every summary is marked as AI-generated and links back to the full transcript.
4. Your responsibility
Use Clarity as a helpful aid, not as an authoritative record. Confirm anything important — such as medical instructions, appointments, amounts or legal matters — with the people involved. Clarity is not a medical device and is not suitable for emergencies. See our Terms of Use.
5. Accessibility
Clarity is designed for people who are deaf or hard of hearing, older adults and people with low vision: large adjustable text, adjustable reading speed, high-contrast colors, speaker-change vibration, and support for iPhone accessibility features such as VoiceOver and larger text.
6. Contact
Questions or concerns about Clarity’s AI: admin@artmento.org.