7. Voice Commander Stack

The jetracer_voice package elevates the JetRacer platform into a fully interactive robotic agent, providing Natural Language actuation pathways completely offline.

Why Offline?

Many robot frameworks lazily employ cloud transcription APIs (e.g., Google STT or AWS). If your local Internet cuts out, your robot is paralyzed.

Instead, we employ Vosk. Vosk utilizes tiny quantized acoustic and language models that execute locally on the Jetson Nano CPU, mapping raw incoming audio signals to phonemes and grammar graphs without ever broadcasting your local voice arrays externally.

Architecture

graph TD
    subgraph Audio Frontend
      MIC[USB Microphone] -->|ALSA/PortAudio| DOCK[Docker /dev/snd]
      DOCK --> PY[PyAudio C-Binding Loop]
    end

    subgraph Vosk ML Engine
      PY -->|ByteStream| INF[Kaldi Recognizer]
      INF -->|Transcribes| STR["kitchen"]
    end

    subgraph Navigation Backdoor
      STR -->|Regex Match| HNDL[Voice Commander Node]
      HNDL -->|geometry_msgs/PoseStamped| NAV[(Nav2 Goal Topic)]
    end

How Delivery Works

Inside voice_commander.py, we execute a high-cycle polling loop. 1. PyAudio extracts 4,000-byte buffers directly from /dev/snd. 2. The buffer mathematically passes through KaldiRecognizer. 3. If the resulting JSON string possesses the word kitchen, the Thread locks immediately. 4. It reads the internal $X$, $Y$ destination vectors hardcoded out of main_config.yaml. 5. It crafts a formal PoseStamped message indicating exactly where the car needs to go inside its 2D Map coordinate system. 6. It formally dictates this target structure over the /goal_pose network. Nav2 receives it seamlessly, plotting a trajectory!

---

[!TIP] Next Step: End your journey logically at 08. Central Configuration to tweak all the values making this possible.