demystify-voice

Interruption-aware voice-agent runtime: low-latency barge-in (client-audible stop within one sub-chunk), spoken-prefix truncation (the interrupted assistant turn keeps only the audio the caller actually heard), backchannel vs barge-in discrimination, and multi-turn context repair.

This is the packaged runtime of the dmstfy-voice-core module. The full module docs (architecture, harness, runbooks, contracts) live in the repository at modules/dmstfy-voice-core/.

Install

pip install demystify-voice                 # core runtime + pipeline
pip install "demystify-voice[server]"       # + FastAPI/uvicorn control plane (demystify_voice.api)
pip install "demystify-voice[pg]"           # + Postgres write-behind store adapter

The core install is intentionally light. demystify_voice.api (the HTTP/WS control plane) requires the server extra; the module imports fine without it and raises a clear error if you import the api without the extra.

Public protocol surface

Plugins (e.g. demystify-voice-lingua) code against these stable exports — see docs/contracts/voice-plugin-protocol.md:

• demystify_voice.frames — frame dataclasses
• demystify_voice.stage — Stage, StageContext, Clock
• demystify_voice.pipeline — Pipeline plugin API (insert_stage, replace_stage)
• demystify_voice.PROTOCOL_VERSION — plugin-protocol contract version (MAJOR.MINOR)

Quickstart (offline, zero keys)

from demystify_voice import SessionRuntime  # inject mock providers; drive with the SimClock

MIT licensed.