Performance Tuning

Yumi is optimized to run smoothly entirely on standard home CPUs, requiring no heavy discrete graphics cards (GPUs). To achieve sub-second response times, apply these performance tuning parameters.

---

⚡ 1. Accelerating Speech-to-Text (STT)

Local transcription is typically the heaviest computational step in Yumi's sensory loop.

A. Choose the Right Model Size

To change model scales, run yumi --config and choose Local Whisper:

• tiny: Takes ~200ms to transcribe. Best for budget notebooks or single-board computers (Raspberry Pi 5).
• base: Takes ~500ms to transcribe. Highly accurate, handles standard phrasing easily. Recommended default.

B. Adjust CPU Core Threads

By default, faster-whisper dynamically selects the number of CPU threads based on your hardware configuration. If your CPU has limited cores, or is running hot, you can throttle thread counts in src/yumi/audio/stt.py to prevent CPU resource locks:

# In stt.py initialization
self.model = WhisperModel(model_size, device="cpu", compute_type="int8", cpu_threads=4)

---

🧠 2. Optimizing LLM Response Rates

A. Use Groq Cloud

Groq's LPU hardware delivers tokens at speeds above 200 tokens/sec. We highly recommend using Groq as your default Mind provider for an instantaneous dialogue pacing.

B. Keep Prompts Lean

The size of your system prompt and the length of your active chat history can degrade response times on cloud APIs.

• Prompt design: Keep personality .txt files under 2,000 characters.
• History length: Periodically clear or truncate your message buffer if you experience slow performance.

---

🗣️ 3. Audio Client Echo Cancellation

To prevent Yumi from "hearing herself" talk through your laptop speakers (which would trigger the local VAD, interrupt her voice, and launch a feedback loop), Yumi implements strict audio constraints on the client mic:

// WebUI microphone configuration in index.html
const stream = await navigator.mediaDevices.getUserMedia({
    audio: {
        echoCancellation: true,   // Filters out playing speaker audio!
        noiseSuppression: true,   // Dampens fans and environment hums!
        autoGainControl: true
    }
});

Keep echo cancellation enabled, especially when running on laptops without headphones!

Proceed to the CLI Command Registry to see all keyboard control routes!