VoiceBridge

A multi-platform media system built around the voice — dictation and speech, many models behind one conversation, rooms with several interlocutors, accumulated material and cloned voices, all on owned hardware.

The problem. Voice is the most natural interface to a model and the least sendable data you own. Every part of it — the dictation, the corpus that accumulates, the cloned voice, the history of a room — is a plausible cloud feature and an implausible thing to hand over. And treating voice as a transcription utility misses what it actually is: the place where a conversation with a system becomes a conversation.

What runs. A server on a consumer GPU with desktop and Android clients. Speech in: streaming transcription with voice activity detection and per-utterance language detection, held on the GPU and unloaded when idle so the card stays free. Speech out: a router across cloud and self-hosted engines, including a locally hosted clone of a specific voice. Around it, the parts that make it a system rather than a feature — an accumulating media library, personas and profiles per participant, memory that survives a session, several models addressable inside one conversation, and rooms where more than one of them sits at the table. Around 9.5k lines across Python and Kotlin.

Designed next. Rooms as an instrument rather than a convenience: asymmetric information between seats, structured history that survives a participant, payoffs that actually differ. The research lab exists to answer exactly that question, and its answers land back here.

The transferable part. What stays local is one decision, made once, at the boundary — and it decides the product, not just the deployment. The corpus, the clones and the room histories live on hardware their owner controls, which is why this repository is the one that stays closed.