Aloud — verbatim relay with a receipt

A deaf, hard-of-hearing or non-speaking person places an ordinary phone call. They type; their words are spoken aloud on the line, unchanged. The other party talks; it is captioned live. After every exchange, the screen shows what was typed next to what the provider's own system recorded as having been said. When the call ends, the recording the provider made is deleted, and the person watches it happen. Type-to-speak with live captions on a phone call is not a new idea — it already ships in the United States, funded by the FCC. InnoCaption added an AI text-to-speech suite in March 2026, and Rylo, an FCC-certified relay provider, lets a caller "speak or type to reply." Anyone pitching this as an empty category is wrong, and this project does not make that claim. What those products do differently is the gap this one is built to close. InnoCaption's own "AI Refine" feature turns a few typed words into "complete, natural-sounding sentences" chosen by a model, not by the caller — and federal relay regulation (47 CFR § 64.604(a)(2)(ii)) requires that a relay carry a conversation verbatim. Aloud registers its own endpoint as the call's language model and, in verbatim mode, that endpoint runs no inference at all — it echoes exactly the characters typed. There is nothing to catch a model rewriting the caller's words in verbatim mode, because there is no model in that path to do it. The transcript AssemblyAI records of what was actually spoken becomes a receipt the caller can check line by line. The same regulation forbids a relay from keeping a call's content beyond the call. Every provider asserts it complies. Aloud calls the provider's own delete endpoint on hangup and shows the confirmation on screen, instead of asking the caller to take retention on faith. Built on the AssemblyAI Voice Agent API for the AssemblyAI Voice Agent Hackathon (lablab.ai, September 2026): a browser-to-browser demo, one WebSocket, no third person listening in.