Voice Assistant Example

A fully offline voice assistant that listens for a wake phrase, transcribes speech, generates an LLM response, and speaks the answer aloud — no cloud required.

Features

• ✅ Wake phrase detection ("hey kdeps") using whisper offline
• ✅ Continuous audio capture from a microphone
• ✅ Offline speech-to-text transcription
• ✅ LLM response generation with Ollama (llama3.2:1b)
• ✅ Offline text-to-speech output using Piper TTS
• ✅ Fully air-gapped — no internet connection needed

Prerequisites

Install required tools:

# Piper TTS (for speech output)
pip install piper-tts

# whisper (for wake detection + transcription)
pip install whisper

# Ollama (for LLM)
curl -fsSL https://ollama.ai/install.sh | sh
ollama pull llama3.2:1b

Configure Your Microphone

Find your audio device:

# Linux
arecord -l

# macOS / Windows
ffmpeg -list_devices true -f avfoundation -i dummy   # macOS
ffmpeg -list_devices true -f dshow -i dummy          # Windows

Update workflow.yaml with your device identifier:

audio:
  device: hw:0,0          # Linux example
  # device: "Built-in Microphone"   # macOS
  # device: "Microphone (Realtek Audio)"  # Windows

Run

# From examples/voice-assistant directory
kdeps run workflow.yaml

# Or from project root
kdeps run examples/voice-assistant/workflow.yaml

Once running, say "hey kdeps" followed by your question. The assistant will:

1. Detect the wake phrase
2. Capture and transcribe your speech
3. Generate an LLM response
4. Speak the response aloud

Structure

voice-assistant/
├── workflow.yaml              # Input sources, activation, and transcription config
├── components/
│   └── tts/
│       └── component.yaml     # .komponent: offline TTS via espeak
└── resources/
    └── respond.yaml           # LLM chat resource (takes transcribed speech as input)

The tts component encapsulates the run.tts executor and is auto-loaded from the components/ directory. Swap the TTS engine by replacing the component without touching the workflow:

# Build a packaged version to share or version-control
kdeps package components/tts --output components/

# Or install a community component
kdeps component install tts

How It Works

Pipeline

Microphone → Wake Phrase Detection → Audio Capture → Transcription → LLM → TTS → Speaker

Activation Loop

The runtime captures chunkSeconds (3s) of audio in a continuous loop. Each chunk is transcribed with whisper tiny and checked against the wake phrase "hey kdeps" with 90% similarity threshold. When detected, the full workflow runs.

Key Expressions

Expression	Description
inputTranscript	The transcribed speech text
get('respond')	The LLM response text
ttsOutput	Path to the generated audio file

Customization

Use a Different LLM

# workflow.yaml
agentSettings:
  models:
    - mistral:7b             # or any Ollama model

# resources/respond.yaml
chat:
  model: mistral:7b

Change the Wake Phrase

activation:
  phrase: "okay computer"
  sensitivity: 0.85          # Lower = more flexible matching

Use eSpeak Instead of Piper (Lighter Weight)

Edit components/tts/component.yaml:

tts:
  mode: offline
  offline:
    engine: espeak
    voice: en

Raspberry Pi / Edge Device Optimizations

For resource-constrained hardware, use the smallest models:

activation:
  offline:
    engine: whisper
    model: tiny              # Smallest, fastest

transcriber:
  offline:
    engine: whisper
    model: tiny              # Use tiny on Pi Zero / limited RAM

# resources/respond.yaml
chat:
  model: llama3.2:1b         # Smallest Llama model

# components/tts/component.yaml
tts:
  offline:
    engine: espeak           # Lightest TTS engine

See Also

• Input Sources Documentation
• TTS Resource Documentation
• Telephony Bot Example — Cloud-based call handler
• Video Analysis Example — Camera surveillance + AI