Vosk Offline STT

Setup

~/.agentos/models/vosk/

modelPath

providerOptions

# Example: download the small English model
wget https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip
unzip vosk-model-small-en-us-0.15.zip -d ~/.agentos/models/vosk/

Configuration

{
  "voice": {
    "stt": "vosk"
  }
}

{
  "voice": {
    "stt": "vosk",
    "providerOptions": {
      "modelPath": "/opt/models/vosk-model-en"
    }
  }
}

Provider Rules

• Audio input must be 16 kHz LINEAR16 PCM. Resample other formats before streaming.
• Model quality scales with model size. Use

  vosk-model-en-us-0.22

  for best English accuracy; use small models on constrained hardware.
• No API key required. The only requirement is a pre-downloaded model directory.
• Streaming recognition is natively supported by the Vosk recognizer.

Examples

• "Use offline Vosk STT for this air-gapped deployment."
• "Transcribe my voice locally without sending audio to the cloud."
• "Configure Vosk with the large English model for best accuracy."

Constraints

• Requires the Vosk npm package and a pre-downloaded model directory.
• Accuracy is lower than cloud providers, especially for accented speech and domain-specific vocabulary.
• Audio must be 16 kHz mono LINEAR16 PCM. Other sample rates or formats require conversion.
• Model download size ranges from ~40 MB (small) to ~1.8 GB (large en-us).