/// Project · ML / WAKE WORD
Giarvis
A custom wake word, "Ei Giarvis", trained for an Italian voice assistant: 94% detection at 2 false alarms an hour.
notes
Off-the-shelf voice assistants accept a handful of English wake words. This one answers to "Ei Giarvis", an Italian take on Jarvis, through a model trained for that phrase alone.
The model is 813 KB of ONNX on top of the openWakeWord feature pipeline. Measured on nine hours of real audio, it detects the phrase 94% of the time at two false alarms per hour.
The first model did not work, but it taught me something: trained on samples that were too clean, it had learned to tell synthetic voices from real ones instead of recognising the word. The problem was in the data, not the network.
An Android app runs the same chain on a wall tablet, verified step by step against the Python reference. Detection works on the device; command transcription is the next step.