Problem
A fast voice loop can become unsafe when every stage acts implicitly.
Capture, remote transcription, drafting, and playback have different privacy and failure modes. Coupling them hides what left the device and why audio played.
Independent build · Windows voice tooling
A voice-practice bridge with one local owner and two explicit provider lanes.
Bounded or realtime capture, Nova-3 transcription, verified role-pack answers, optional grounded drafting, and deliberate SAPI or Aura playback—each kept as a visible action.
Build
The local answer path remains useful on its own.
I separated bounded and realtime capture, Nova-3 transcription, deterministic role-pack ranking, optional grounded Codex drafting, human review, and dual-device Windows playback.
Result
Remote services can fail without taking the whole session with them.
Local matching stays available, drafting fails closed at its latency gate, SAPI remains the default, Aura testing stays explicit, and metadata-only logging excludes content and secrets.
Before / after
Keep the local answer path available.
Broken
answer = remote_draft(question)
play(answer)Fixed
answer = role_pack.best_match(question)
if operator_requests_draft:
answer = grounded_draft(question, timeout=limit)
play_when_confirmed(answer, voice=selected_voice)Evidence
The bounded path worked under controlled practice conditions.
- Synthetic room-audio exercises completed the intended question → answer → playback loop.
- A Windows executable and portable ZIP were built and inspected historically.
- The private-project boundary excludes persistent audio and transcript storage.
Stack and capabilities
A portable Windows audio system with visible boundaries.
- Python
- Deepgram Nova-3
- Deepgram Aura-2
- Windows SAPI
- Role packs
- Realtime session states
- Native dual-output playback
- PyInstaller
Next