Sophon-Realtime marked version 1.3.0 on November 14 with reference-voice support in its E2E build and a Mac startup fix for some audio devices. This is a specific update to an open voice-AI project. It is not a newly trained model family: the changelog explicitly says that weight files have not changed.
What changes, and for whom
The E2E addition uses a short reference recording without further training. The maintainers restrict its intended use to a person's own voice or one they have permission to use. The change applies to the vLLM-Omni E2E implementation; the separate custom engine continues to use its built-in voice. A shared version number therefore does not imply identical features in every deployment path.
The server guide describes speech-to-speech processing and realtime conversation interfaces. For E2E, reference audio is prepared and cached; an unreadable or overly short recording falls back to the built-in voice. That detail matters during evaluation: hearing an answer does not establish that the intended configuration was used. Both the selected implementation and the actual output need to be checked.
The family is presented as full-duplex, meaning that it can listen while speaking and react to interruptions. This describes conversation handling, rather than guaranteeing factual answers or correct interpretation of every interruption. Published performance measurements come from the developers. SIExpo has not reproduced those tests and does not treat them as independent validation.
The local Mac path has separate requirements
The Mac guide specifies Apple silicon, macOS 14 or later and recommends at least 36 GB of unified memory. It also describes operation with 32 GB when other applications are closed; that is not a promise of the same headroom under every workload. Approximately 30 GB of free disk space is required. A machine without a microphone needs an external audio device.
The November 14 fix addresses a startup problem with Bluetooth, AirPlay and HDMI outputs involving large delays in echo cancellation. The announcement describes tests with simulated echo. This establishes a developer-reported check, rather than compatibility with every headset, television or audio driver. Because the documentation continues to change, retaining the exact revision is sensible for a repeatable trial.
How to assess practical value
Our editorial interpretation is that local execution and configurable voices give developers more options while leaving device and conversation validation in their hands. A small pilot should compare the transcript with what was actually spoken, include interruptions and background noise, and record memory use. Fictional test conversations and an authorized personal voice provide a useful starting point.
This release does not establish reliable Bulgarian-language support or readiness for every customer-service line. Open code makes inspection and repetition easier, but language quality, recording protection and application reliability remain separate questions. A local model also needs an application that handles microphone access and any optional network connections deliberately; model location alone does not describe the entire data path.
The cover is an AI-generated editorial illustration, not a documentary photograph.



