Low Latency and Efficient Voice Activator for Edge Devices

Background As voice-controlled IoT proliferate, processing everything in the cloud is too costly, privacy-invasive, and slow. The future belongs to hybrid architectures where the edge handles the initial 'wake-up' and the cloud handles the heavy lifting.

Description Build an ultra-lightweight, highly accurate keyword spotting (KWS) model that runs locally on a low-power device. Upon detecting the keyword, the system must instantly and efficiently stream the subsequent audio to a remote Automated Speech Recognition (ASR) server with minimal data overhead and latency.

Key Metrics for Evaluation

• Efficiency: Model size (RAM/Flash footprint) and CPU usage during idle listening.
• Accuracy: High true-positive rate for the keyword with near-zero false activations.
• Latency: The time delta between the keyword ending and the cloud ASR receiving the audio stream.

Software & Framework Restrictions

• Open-Source Only: The use of proprietary, closed-source, or commercial voice-activation SDKs is strictly prohibited.
• Allowed Frameworks: Teams must build their keyword spotting (KWS) pipelines using open-source machine learning and TinyML frameworks. Recommended tools include TensorFlow Lite for Microcontrollers, PyTorch Mobile or similar.
• No Pre-Trained Global Keywords: Teams cannot use models pre-trained on generic smart-assistant keywords like 'Hey Google' or 'Alexa'. They need to train on a custom key word.

Expected Solution Teams are expected to deliver a robust, deployable system architecture. A successful submission must strictly satisfy the following technical boundaries:

• Hardware & Runtime Environment: The edge software application must run smoothly within an environment restricted to less than 256KB of RAM and consume under 10% CPU utilization while idling in continuous listening mode. Heavy or uncompressed pre-trained transformers are disqualified. Solutions will be formally evaluated on physical low-power microcontrollers (e.g., Raspberry Pi or ESP32).
• Model should work for the given custom key word.