AI Engineer for Custom OpenWakeWord Project

Train OpenWakeWord with a Custom Wake Word (High Accuracy, Low False Positives)

Project Overview

I am looking for an experienced AI/ML engineer to train OpenWakeWord with a custom wake word. The goal is to build a production-quality wake word model that is highly accurate, responsive, and resistant to false activations.

Scope of Work

* Train a custom OpenWakeWord model from scratch or fine-tune an existing pipeline.
* Use real positive voice samples from multiple speakers, accents, speaking speeds, distances, and environments.
* Collect and use a large negative dataset containing similar-sounding words, random conversations, TV/audio, background noise, music, and other speech to minimize false positives.
* Optimize the model for:
* High detection accuracy
* Extremely low false positive rate
* Fast response time
* Robust performance in noisy environments
* Test extensively with real-world audio and provide performance metrics.
* Deliver the trained model along with inference files and clear integration instructions.

Required Skills

* OpenWakeWord
* Machine Learning / Deep Learning
* Speech Recognition / Audio Processing
* Python
* ONNX / TFLite (preferred)
* Experience training keyword spotting (KWS) or wake-word models

Deliverables

* Fully trained custom OpenWakeWord model.
* Training scripts/configuration (if applicable).
* Test results showing detection accuracy and false positive performance.
* Documentation explaining how to use and retrain the model if needed.

Proposal Requirements

Please include:

* Previous experience with OpenWakeWord or custom wake-word models.
* Relevant AI/audio processing projects.
* Expected timeline.
* Fixed-price quote.
* Any suggestions for improving wake-word reliability and reducing false activations.

Important: The final wake word should be highly reliable, with minimal false positives and strong performance across different speakers and environments.
** Training through opencollab is not acceptable.