The ggml‑org/llama.cpp project released version b11780 on Oct 21 2026. It introduces a realtime transcription API that runs over websockets. The API checks an API key during the websocket handshake and returns 401 if the key is missing or wrong. Audio streams are now started as soon as the server slot is live, preventing dead‑locks that previously held the session open. The implementation follows the Google‑Assistant‑style Realtime schema, so clients must send audio at the model’s native sample rate.
Why it matters
Developers can now add live speech‑to‑text to apps using llama.cpp without building their own server infrastructure.
