Local Concierge Lab

Core stack

Component	Model / Tool	Runs locally?
Automatic speech recognition	openai/whisper-medium.en (Hugging Face transformers)	Yes
Language model	llama3:latest via Ollama	Yes
Text-to-speech	Piper voice en_US-lessac-medium	Yes

Setup

1. Create/activate a Python environment (we used conda, but venv works too).
2. Install dependencies:

   pip install -r requirements.txt

3. Install Ollama, run ollama serve (or ./start_ollama.sh), and pull the model:

   ollama pull llama3

Running the pipeline

Only four CLI arguments are provided to keep things simple. Run the main script like so:

python main.py \
  --audio-path data/sample_call.wav \
  --chunk-duration 2.0 \
  --chunk-overlap 0.5 \
  --scenario restaurant

• --audio-path: the caller recording we want to process.
• --chunk-duration: seconds of audio per ASR window to simulate real time chunking.
• --chunk-overlap: seconds of context we carry between windows.
• --scenario: persona prompt (restaurant, mental, coach).

The script reads the audio, prints each Whisper chunk, generates a single reply with the selected persona, and saves the Piper output to artifacts/agent-response.wav.