Five separate systems. One badge. Here's how we orchestrated wake word detection, audio embeddings, LLM inference, and TTS into a seamless voice-activated experience-all within 8MB of RAM.
When Speech-to-Text models were too big for our 8MB PSRAM, we pivoted to a clever alternative: contrastive learning to match spoken audio directly to pre-embedded intents. Here's how we trained that system.
How we trained a custom 'Hey Daisy' wake word detector using confusable negatives, synthetic voice generation, and deployed it on ESP32-S3 with EdgeNeuron TFLite.
Deep dive into running a Large Language Model on an ESP32-S3 microcontroller - exploring llama2.c, SIMD optimizations, and the challenges of streaming inference on embedded hardware.
