Skip to content

OpenWakeWord

Overview

OpenWakeWordNode is a ROS 2 node that performs keyword spotting (wake word detection) using models trained with OpenWakeWord. It supports inference using ONNX or TensorFlow Lite (TFLite) frameworks. The node subscribes to raw audio data, detects wake words in real time, and publishes detection events.

What is OpenWakeWord?

OpenWakeWord is an open-source wake word detection toolkit that enables training custom wake word models (e.g., "frida", "no", "stop", "yes") and running inference efficiently. It supports exporting models to ONNX and TFLite formats for deployment on various platforms.

Use Case

This node is intended to detect wake words during the @Home competition to trigger specific functions

Architecture and Implementation

ROS Parameters

Parameter Type Default Path / Value Description
model_path string /workspace/src/hri/packages/speech/assets/oww Directory containing ONNX wake word models (.onnx files)
inference_framework string onnx Framework used for inference (onnx or tflite)
audio_topic string /rawAudioChunk Topic where raw audio chunks are received
WAKEWORD_TOPIC string /speech/oww Topic to publish detected wake words
chunk_size int 1280 Number of samples processed per inference
detection_cooldown float 1.0 Minimum time (in seconds) between detections to avoid repeated triggers
SENSITIVITY_THRESHOLD float 0.5 Confidence threshold required for detection

Model Init

Feature Description
Load Wakeword Models Loads all .onnx models from model_path. Logs number of models loaded.
Default Model Fallback Loads default wake word model if no directory or models are found.
Feature Extractor Download Downloads melspectrogram.onnx and embedding_model.onnx if missing.
Path Setup Copies models to the correct execution path if not already present.
ROS Publisher/Subscriber Creates publisher (WAKEWORD_TOPIC) and subscriber (audio_topic).
Cooldown Timer Initializes internal timer to prevent repeated wake word detections.

Audio Inference Flow

Step Description
Receive Audio Listens to AudioData messages from audio_topic.
Convert to NumPy Converts audio byte stream to np.int16 NumPy array.
Run Inference Feeds the audio to OpenWakeWord model for prediction.
Monitor Prediction Buffer Checks latest confidence score for each keyword.
Detection Condition If score > SENSITIVITY_THRESHOLD and cooldown has passed, a detection is triggered.
Publish Detection Sends a std_msgs/String message like {"keyword": "frida", "score": 0.87}.

Parameters

Parameter Type Default Description
model_path string /workspace/src/hri/packages/speech/assets/oww Directory containing wake word .onnx model files
inference_framework string onnx Framework used for model inference (onnx or tflite)
audio_topic string /rawAudioChunk ROS topic subscribing to raw microphone audio data
WAKEWORD_TOPIC string /speech/oww ROS topic publishing detected wake word messages
chunk_size int 1280 Size of audio chunks processed per prediction
detection_cooldown float 1.0 Minimum seconds between repeated detections
SENSITIVITY_THRESHOLD float 0.5 Confidence score threshold to confirm wake word detection

Message Types

  • Input: frida_interfaces/msg/AudioData — raw audio buffer (16-bit PCM)
  • Output: std_msgs/msg/String — JSON string with detected keyword and confidence score, e.g. {"keyword": "frida", "score": 0.87}