> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Audio Assistant

> Voice Activity Detection (VA) and acoustic event detection (SSG) for edge audio workflows.

The **Audio Assistant** node sits between an [Audio Track](/use-cyberwave/workflows/audio-track/overview) trigger and downstream nodes (STT, alerts, models). It normalizes all input to **PCM S16LE int16 @ 16 kHz mono** via the shared audio ingress layer—the same contract as the rest of the audio pipeline.

## Pipeline position

```
Audio Track Trigger → Audio Assistant → Call Model / Send Alert / …
```

## Modalities

| Modality                       | Purpose                                 | Engine             |
| ------------------------------ | --------------------------------------- | ------------------ |
| **Voice Assistant (VA)**       | Segment speech utterances               | Silero VAD         |
| **Sound Security Guard (SSG)** | Detect security-related acoustic events | MIT AST (AudioSet) |

Only one modality is active per node instance.

## Shared audio contract

| Direction | Key              | Format                                                                             |
| --------- | ---------------- | ---------------------------------------------------------------------------------- |
| Input     | `audio`          | PCM S16LE numpy `int16`, float32, raw bytes, or WAV—adapted to int16 @ 16 kHz mono |
| Output    | `audio`          | int16 mono @ 16 kHz (when a segment or alert window is emitted)                    |
| Output    | `sample_rate_hz` | Always `16000`                                                                     |
| Output    | `channels`       | Always `1`                                                                         |

Upstream **Audio Track** buffer presets control chunk size into the node:

| Preset                  | Chunk size    | Typical use                              |
| ----------------------- | ------------- | ---------------------------------------- |
| Voice Assistant (32 ms) | 512 samples   | VA streaming (matches Silero frame size) |
| Wake Word (80 ms)       | 1280 samples  | Wake Word Engine                         |
| Speech-To-Text (4 s)    | 64000 samples | Batch / long windows                     |

For **SSG**, prefer accumulating at least **1–4 s** per analysis window (default analysis preset: Speech-To-Text 4 s).

## Edge execution

Audio Assistant nodes are **edge-only**. Workflows compile to a `wf_*.py` worker module and sync to the device:

```bash theme={null}
cyberwave workflow compile <workflow-uuid>    # inspect emitted source + warnings
cyberwave workflow sync --twin-uuid <twin>  # deploy to edge
```

Typical chain:

```
Audio Track (@cw.on_audio) → Audio Assistant → Send Alert / Call Model
```

Compile checks:

* **VA** upstream Audio Track must use buffer preset **Voice Assistant (32 ms)**.
* **SSG** upstream Audio Track should use **Speech-To-Text (4 s)** (or custom ≥1 s).

The compiler emits a warning listing required Python extras when VA/SSG nodes are present.

## Edge dependencies

```bash theme={null}
pip install "cyberwave[ml-vad]"      # Voice Assistant (Silero)
pip install "cyberwave[ml-aed]"      # Sound Security Guard (transformers + AST)
```

The **edge-ml-worker** container image pre-installs `zenoh`, `ml-vad`, `ml-aed`, and `ml-wakeword`, and pre-downloads AST and OpenWakeWord weights for air-gapped use.

SSG on bare-metal edges downloads `MIT/ast-finetuned-audioset-10-10-0.4593` on first run (\~340 MB) unless baked into the image.

## Pages

* [Voice Assistant (VAD)](/use-cyberwave/workflows/audio-assistant/voice-assistant)
* [Sound Security Guard (AED)](/use-cyberwave/workflows/audio-assistant/sound-security-guard)
* [Technical reference](/use-cyberwave/workflows/audio-assistant/technical-reference)