kawaiipunk
bc02be00cd
- Resample PCM in the audio worklet when the browser cannot open a 16 kHz AudioContext (Firefox), keeping the native 16 kHz path where available; add Node coverage in tests/resampler.test.cjs. - Default the capture page relay and caption-site fields to the self-hosted relay/UI, and only let a ?relay= link override the prefilled address. - Remember the service access token in sessionStorage for the tab session so the redirect target does not require re-entry on every load. - Add an inline favicon to stop the token-protected /favicon.ico 401. - Drop the invalid ws://[::1]:* Content-Security-Policy source. - Document Chrome/Edge/Firefox capture.
Guides and help
Start with one microphone on one computer. Choose a setup, get your first captions working, then add an editor, audience displays or more streams.
New users
- Visual introduction: workflow diagrams, the actual capture screen, hosting choices and what you will need.
- Install and get your first captions: Windows, Linux and Docker commands; connect, choose a microphone, capture, save and stop.
- Troubleshooting: connection errors, microphone permission, missing captions and growing delays.
Choose your setup
- Hardware and quality profiles: CPU, NVIDIA, private VPS, model accuracy and measured concurrent-stream limits.
- Windows and NVIDIA setup: tested requirements, exact commands and GPU verification.
- Installation and deployment reference: native Python, Docker, model downloads and remote access.
Connect an editor or audience
- Caption.ninja integration: the separate capture page, public/private sharing, direct overlays and hosted-page permissions.
- Run your own caption relay: Windows/Linux setup, room tokens, editor/viewer connections, OBS links and TLS.
- First-event checklist: rehearse microphones, caption review, audience display, failures and shutdown together.
Operate or integrate
- Operations: supervision, logs, offline operation, upgrades and recovery.
- API reference: native live-audio requests and the supported OpenAI-style WAV subset.
- Security: trusted-producer access, tokens and reporting issues.
- Optional AI setup assistance: a setup prompt and deployment skill.
- Contributing: development and regression checks.
Read the measurements
- CPU and language results
- NVIDIA hour and limitations
- Text relay capacity and recovery
- Page layout and navigation review
- Full regression and clean-install checks
Speech recognition and caption delivery have different limits. A relay benchmark does not establish how many microphones your inference computer can transcribe.