WaveCap Configuration Skill

Use this skill to view and modify WaveCap configuration settings.

Configuration File Location

The main configuration file is at:

/Users/thw/Projects/WaveCap/state/config.yaml

View Current Configuration

Full Config

cat /Users/thw/Projects/WaveCap/state/config.yaml

Whisper Settings

cat /Users/thw/Projects/WaveCap/state/config.yaml | grep -A 30 "^whisper:"

LLM Correction Settings

cat /Users/thw/Projects/WaveCap/state/config.yaml | grep -A 15 "^llm:"

Stream Definitions

cat /Users/thw/Projects/WaveCap/state/config.yaml | grep -A 100 "^streams:"

Key Configuration Sections

Server Settings

server:
  host: 0.0.0.0
  port: 8000
  corsOrigin: "*"

Whisper Transcription Settings

whisper:
  model: large-v3-turbo          # Model: base.en, small, medium, large-v3-turbo
  language: en                    # Language code
  chunkLength: 20                 # Seconds per chunk (15-120)
  minChunkDurationSeconds: 12     # Minimum chunk before flush
  contextSeconds: 0.5             # Overlap for context
  silenceThreshold: 0.02          # Audio energy threshold
  silenceHoldSeconds: 1.2         # Wait after silence
  maxConcurrentProcesses: 2       # Parallel transcriptions
  beamSize: 8                     # Decoder beam width
  decodeTemperature: 0.0          # 0.0 = deterministic

LLM Correction Settings

llm:
  enabled: true                   # Enable/disable LLM correction
  model: llama-3.2-3b             # Model name (see available models)
  temperature: 0.1                # Generation temperature
  maxTokens: 256                  # Max output tokens
  minTextLength: 10               # Skip short texts
  domainTerms:                    # Custom terminology
    - SITREP
    - SAPOL

Available LLM Models

• llama-3.2-1b - Fast, small (Llama-3.2-1B-Instruct-4bit)
• llama-3.2-3b - Balanced (Llama-3.2-3B-Instruct-4bit)
• qwen-2.5-1.5b - Fast alternative (Qwen2.5-1.5B-Instruct-4bit)
• qwen-2.5-3b - Balanced alternative (Qwen2.5-3B-Instruct-4bit)
• llama-3.2-8b - Higher quality (Llama-3.2-8B-Instruct-4bit)
• qwen-2.5-7b - Higher quality alternative
• deepseek-r1-8b - DeepSeek R1 distilled

UI Settings

ui:
  themeMode: system               # system, light, dark
  colorCodingEnabled: false       # Color by confidence
  transcriptCorrectionEnabled: false

Stream Configuration

streams:
  - id: unique-stream-id
    name: "Display Name"
    url: https://stream-url       # For audio streams
    source: audio                 # audio, pager, remote, combined
    enabled: true
    pinned: false
    ignoreFirstSeconds: 30        # Skip initial audio
    recordingRetentionSeconds: 604800  # 7 days

Common Configuration Changes

Change Whisper Model

Edit config.yaml and change:

whisper:
  model: small  # or base.en, medium, large-v3-turbo

Then restart the backend.

Adjust Transcription Latency

For faster updates (shorter chunks):

whisper:
  chunkLength: 15
  minChunkDurationSeconds: 8

For better sentence structure (longer chunks):

whisper:
  chunkLength: 45
  minChunkDurationSeconds: 30

Enable/Disable LLM Correction

llm:
  enabled: false  # or true

Add Domain Terms for LLM

llm:
  domainTerms:
    - SITREP
    - SAPOL
    - YOUR_CUSTOM_TERM

Add Silence Hallucination Phrases

Phrases to filter when detected during silence:

whisper:
  silenceHallucinationPhrases:
    - "thank you"
    - "transcription by castingwords"

Add a New Stream

Add to the streams array:

streams:
  - id: my-new-stream
    name: "My New Stream"
    url: https://example.com/stream
    source: audio
    enabled: true

After Configuration Changes

Most config changes require a backend restart:

# Kill existing backend
pkill -f "uvicorn.*wavecap"

# Restart
cd /Users/thw/Projects/WaveCap/backend
source .venv/bin/activate
uvicorn wavecap_backend.server:create_app --factory --host 0.0.0.0 --port 8000

View Runtime Configuration via API

UI configuration (subset):

curl -s http://localhost:8000/api/ui-config | jq

Logging configuration:

curl -s http://localhost:8000/api/logging-config | jq

Access/auth configuration:

curl -s http://localhost:8000/api/access | jq