| --- |
| license: other |
| private: true |
| tags: |
| - vigil |
| - wake-word |
| - inference-artifacts |
| - ros |
| - qwen-asr |
| --- |
| |
| # VIGIL Trigger Inference Artifacts |
|
|
| Private integration bundle for the VIGIL two-stage trigger used by the ROS `medical_agent` integration. |
|
|
| This repository contains only the small VIGIL inference heads and public configuration needed for integration testing. It does not contain raw audio, feature caches, predictions, datasets, Qwen weights, openWakeWord weights, logs, or private participant metadata. |
|
|
| ## Contents |
|
|
| ```text |
| PUBLIC_MANIFEST.json |
| stage1/checkpoint_best.pt |
| stage1/model_config.json |
| stage1/threshold.json |
| stage1/metrics.json |
| stage2_bce_supcon/checkpoint_best.pt |
| stage2_bce_supcon/threshold.json |
| stage2_bce_supcon/metrics.json |
| stage2_bce_supcon/frozen_qwen_integrity.json |
| ``` |
|
|
| ## Selected Runtime Configuration |
|
|
| - Stage 1 feature backend: official openWakeWord speech embedding backbone. |
| - Stage 1 trainable head: LayerNorm -> 2-layer GRU -> Linear. |
| - Stage 1 threshold: `0.9973222613334656`. |
| - Stage 2 variant: `stage2_bce_supcon`. |
| - Stage 2 input: frozen Qwen audio encoder features. |
| - Stage 2 threshold: `0.9877771735191345`. |
| - Qwen model: load separately from `Qwen/Qwen3-ASR-1.7B`. |
| - Qwen trainable parameters: zero. |
|
|
| ## Intended Integration Shape |
|
|
| The ROS node should run raw audio in parallel with ASR: |
|
|
| ```text |
| 16 kHz microphone audio |
| -> rolling windows |
| -> Stage 1 candidate detector |
| -> Stage 2 verifier |
| -> VIGIL trigger event |
| |
| ASR transcript |
| -> command/question text after trigger |
| ``` |
|
|
| After VIGIL is confirmed, `medical_agent` can classify the transcript as a command or question and route it to command handling, VQA, or the medical LLM. |
|
|
| ## Safety Notes |
|
|
| - Do not treat this repository as a training dataset. |
| - Do not publish publicly until the VIGIL team approves sharing trained voice-trigger heads. |
| - Do not add raw participant audio, feature caches, predictions, exports, or local logs. |
|
|
