llms.txt
# Cadence
> Cadence is single-device live captioning for deaf and hard-of-hearing users. Speech recognition, speaker separation, and punctuation all run on an iPhone. Audio and captions never touch a network, not for processing, not for telemetry.
Made by Index Labs LLC. iPhone only, iOS 17 or later.
## What it does
- **Captions the moment you open it.** The capture screen starts listening on its own after a one-time permission prompt. Pause, resume, and finish are one tap each, with a status dot showing whether it's actually listening.
- **Flip the screen.** One tap turns the captions 180° so someone sitting opposite can read along. It unflips when you stop.
- **Speaker labels.** Who said what, marked as Speaker A, B, C with a coloured gutter bar. Anonymous, and reset every session.
- **Text as large as you need.** A caption size slider scales past twice standard size, on top of your Dynamic Type setting.
- **Scroll back without losing the thread.** New captions stop yanking you down; a "jump to latest" pill returns you.
- **Sessions saved on the phone.** Kept on the phone with date, length, and a preview of the opening line. Shareable as plain text with speaker names.
- **Mic tuning.** Tell it group or one-on-one, and how readily to treat quiet or distant sound as speech.
- **Loud failures.** A stalled pipeline gets a banner and a retry button; a denied mic gets a banner pointing to Settings. Never a frozen screen.
Behind those features are four recognition engines. They are not quality tiers; Turbo and Complex serve two genuinely different sensory situations:
- **Built In** (default, no download, needs iOS 26): Apple's on-device speech recognition. Works the moment you install. No speaker labels.
- **Built In (Legacy)** (no download): Apple's older engine, for the widest device support. No speaker labels.
- **Turbo** (224 MB): Lowest latency, for users with residual hearing or who lipread: the caption has to land near the lip movement. No speaker labels.
- **Complex** (639 MB + 45 MB for labels): Highest accuracy, for profoundly deaf users where the caption *is* the information channel. The only engine with speaker labels.
## Core invariants
These are accessibility requirements, not preferences:
1. **Committed text never reflows.** Once rendered, text changes only in colour or weight, never re-wraps or reorders.
2. **No network with audio, ever.** No cloud fallback, no caption or audio telemetry, no model-improvement upload.
3. **Modes never auto-switch on quality.** The only permitted automatic change is thermal, and it announces itself.
4. **Speaker identity is never colour alone.** Always colour plus a persistent letter plus a stable gutter position.
5. **Failures are loud.** A frozen caption looks like silence; the app says so instead.
6. **VoiceOver works.** Dynamic Type scales to AX5 without breaking layout.
## Privacy summary
No user accounts, no servers, no analytics. Audio is held in memory only long enough to caption it, then discarded, never written to a file, never uploaded. Captions are saved to the app's private local storage for review and can be deleted in-app. Speaker labels are anonymous (A, B, C); no voiceprint or biometric template is stored, and labels reset every session. Optional contacts access (off by default) reads only given names, as a spelling dictionary, never to identify a speaker.
The only network request Cadence makes is downloading speech models from FluidInference (maintainers of the open-source FluidAudio project). That request contains nothing about the user.
## Important limitation
Cadence is not a substitute for a human interpreter. Automatic speech recognition can produce captions that are fluent, confident, and wrong. It is not a certified interpretation service, not a legal record, and does not satisfy any institution's ADA accommodation obligations.
## Pages
- [Home](https://cadencecaptions.com/): Product overview, the four engines, invariants, and FAQ.
- [Privacy Policy](https://cadencecaptions.com/privacy): What is processed on-device and why there is no backend.
- [Terms of Service](https://cadencecaptions.com/terms): Usage terms, interpreter disclaimer, and recording-consent guidance.
## Optional
- [Full text](https://cadencecaptions.com/llms-full.txt): Complete site content, including both legal documents in full.
