\n \n \n \n

Voice systems build · Private practice tool

Fast voice assistance with explicit controls and fail-closed boundaries.

Niyam Voice Bridge combines bounded microphone capture, Deepgram transcription, local deterministic answer matching, optional grounded drafting, and deliberate Windows audio playback.

Stack
Python · Deepgram · Windows audio
Privacy
Practice audio/transcripts kept in memory
Verification
5/5 bounded synthetic room-audio cycles

01 / Design

Nothing should speak or send data by accident.

Capture, transcription, AI drafting, voice generation, and playback are separate explicit actions. Audio and transcripts are not written to repository/runtime logs, and optional remote drafting receives selected verified facts rather than arbitrary files.

02 / Engineering

Keep a fast local path even when remote AI is slow.

  • Push-to-talk physical-microphone capture with bounded duration.
  • Deepgram transcription and deterministic local answer ranking.
  • Optional grounded Codex drafting that fails closed on timeout.
  • Explicit Windows audio routing and dual-output playback work.
  • Metadata-only logging that excludes prompts, answers, transcripts, and secrets.

03 / Evidence boundary

A successful synthetic test is not a product claim.

A five-question synthetic room-audio test completed 5/5 cycles, but the repository explicitly does not claim proven performance for real interviews, arbitrary accents/noise, commercial use, or a passed realtime Codex latency gate.

Back to all work