Virtual voice builder

Wires a real microphone through an AI brain (STT → LLM → TTS) and routes the output to a virtual audio cable so apps like Google Meet hear the processed voice. Use when building a real-time voice interception layer in a Node.js project.

Voice Pipeline Skill

Guide the user through building a real-time audio interception pipeline in Node.js. The pipeline: Mic → ffmpeg resample → Deepgram STT → LLM → Sentence chunker → TTS → PCM decode → VB-Cable.

Capabilities

1. OS Prep: Walk through virtual audio driver install and ffmpeg setup.
2. Device Discovery: Run scripts/01_list_devices.js to find the exact VB-Cable name.
3. Step-by-step build: Execute scripts 01–07 in order, stress-testing each before proceeding.
4. Kill switch: Export a stop() function the host app calls to instantly halt the pipeline.
5. Env wiring: Write required keys to the project's .env file.

Critical Rules

• Never use naudiodon or electron — both are incompatible with a portable Node.js skill.
• Always resample to 16kHz mono PCM before sending to Deepgram.
• Never stream raw LLM tokens to TTS — buffer to sentence boundaries first.
• TTS output (MP3) must be decoded to PCM before writing to VB-Cable.
• The kill switch must be IPC-based, not globalShortcut (Electron-only).

Execution Order

Run each script independently to stress-test before wiring together.

Step 1 → OS prep (no code, see references/step_by_step.md)
Step 2 → scripts/package.json + npm install
Step 3 → node scripts/01_list_devices.js
Step 4 → node scripts/02_capture_resample.js
Step 5 → node scripts/03_deepgram_ws.js
Step 6 → node scripts/04_llm_stream.js + scripts/05_sentence_chunker.js
Step 7 → node scripts/06_tts_ws.js + scripts/07_pcm_write.js

Progressive Loading

For architecture diagram and signal flow: read references/architecture.md. For full corrected step-by-step: read references/step_by_step.md. For all required env vars: read references/env_schema.md.