Skip to main content

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.

The Audio Assistant node 🔊 sits between an Audio Track trigger and downstream STT/model nodes. It uses Silero VAD to intelligently segment continuous audio streams into clean speech utterances before triggering the next pipeline stage.

​
Pipeline Position

Audio Track Trigger → Audio Assistant → Call Model (Whisper/STT)
The Audio Track trigger fires on every raw PCM chunk from a twin’s microphone. Without the Audio Assistant, every chunk — including silence and noise — hits the STT API. The Audio Assistant filters and segments the stream so only meaningful speech reaches downstream models.

​
Modalities

​
🔊 Voice Assistant (VA) — Voice Activity Detection

Uses Silero VAD to detect speech start/end in continuous audio streams. Four pre-configured profiles:
Sub-ModalityIconUse Case
Batch (The Librarian)📖Offline file segmentation for STT. Groups utterances into ~28s chunks for Whisper’s 30s window.
Streaming (The Butler)🎬Real-time command endpointing. Low latency, rejects background noise.
Dictation (The Secretary)🖋️Long-form continuous speech. Tolerates pauses for thinking.
Custom⚙️Full manual control over all Silero VAD parameters.

​
🚨 Sound Security Guard (SSG) — Acoustic Event Detection

Skeleton for future acoustic event classification using Large-Scale Pretrained Audio Neural Networks (PANNs).
Sub-ModalityIconUse Case
Glass Break / Falling Objects💥Detect glass shattering and heavy impacts.
Constant Alarm (Siren)🚨Detect sustained alarm tones and sirens.
”Help!” / Screaming🆘Detect human distress calls and screams.
Custom⚙️Configure custom acoustic event detection.
SSG mode is not yet implemented. Audio will be forwarded unchanged. Runtime will use Large-Scale Pretrained Audio Neural Networks in a future release.

​
Input / Output Format

DirectionKeyTypeDescription
InputaudioAUDIOAccepts PCM S16LE int16 (zero-copy), float32, WAV bytes, or raw PCM bytes. Automatically adapted to int16 @ 16 kHz mono.
OutputaudioAUDIOExtracted speech as PCM S16LE int16 numpy array. Mono, 16 kHz. Only emitted when a complete speech segment is detected (speech end).
Outputspeech_probabilityNUMBERRaw Silero VAD confidence (Python float, 0.0-1.0). Emitted on every chunk — acts as a continuous logic gate / trigger signal.
Outputis_speakingBOOLEANTrue while speech is active. Emitted on every chunk.

​
Two output modes

The Audio Assistant emits outputs on every incoming audio chunk, but the content differs depending on state: While accumulating speech (no complete segment yet): 
{ "speech_probability": 0.87, "is_speaking": true, "sample_rate_hz": 16000, "channels": 1 }
When speech ends (complete segment extracted): 
{
  "audio": "<numpy int16 array>",
  "speech_probability": 0.92,
  "is_speaking": false,
  "sample_rate_hz": 16000,
  "channels": 1,
  "start_timestamp_sec": 1.2,
  "end_timestamp_sec": 3.8
}
The speech_probability output is useful for:
  • Downstream logic gates that need to react before the full segment is ready
  • UI visualizations (waveform activity indicators)
  • Custom nodes that need the VAD trigger signal as a float
See Audio in Workflows for the full audio format specification.

​
Execution Target

Edge-only. VAD requires stateful buffering of a continuous audio stream with sub-100ms latency. This cannot run in the cloud workflow runner.

​
Quick Start

  1. Add an Audio Track trigger node connected to a twin with a microphone
  2. Add an Audio Assistant node downstream
  3. Select Voice Assistant → Streaming (The Butler) for real-time command detection
  4. Connect to a Call Model node with Whisper/STT model
  5. Activate the workflow