Neural Translation Engine™

Sensor Fusion

FLIR Lepton 3.5 provides leaf surface temperature delta. BME688 provides ambient temperature, humidity, pressure, and quad-gas resistance. Time-of-day and season are appended to the feature vector.

Feature Extraction

12 real-time features are normalised and packed into an INT8 input tensor: VPD, leaf temp delta, gas resistance, pressure trend, humidity trend, light proxy, species ID, time-of-day, day-of-year, and 3 rolling averages.

NTE™ Neural Network Inference

The INT8 quantized neural network (ARM Cortex-M4 NPU) runs inference in under 8ms, classifying the input into one of 64 emotional state classes with a confidence score.

Speech Template Selection

The top-confidence emotional state class selects a language-appropriate speech template from a library of 20–40 curated human-written utterances per language per emotional class.

WaveNet-Lite Synthesis

The selected template text is synthesised into 44.1kHz audio using our on-device WaveNet-Lite voice model. Personality modulation (warmth, pitch, cadence) is applied per plant profile.