Arm Virtual Hardware Project Best Practices
Qeexo AutoML’s Arm Virtual Hardware integration supports high sensitivity microphone data collection for training, test, and Live Replay (Live Test) of keyword detection results. High sensitivity data collections may be impacted by small environmental factors, as such we recommend the following when working with Arm Virtual Hardware projects.
Use the same microphone for training, test, or Live Replay data collection
Position yourself directly in front of your computer microphone at a consistent distance when performing data collection and Live Replay
Speak loudly and clearly into the microphone
For training or test data collections, collect at minimum 15 instances for each class; more quality data collected and included in training typically produces better results
For each training or test data collections, each keyword should be segmented and labeled before being used for training or test data evaluation
A baseline class, such as background or silence should be included during training, this can be obtained through segmenting between events or collecting baseline date
Train / build all models and experiment with Live Replay, each model may perform differently depending on the complexity of the problem - typically solid performing models include GBM, RF, XGB, and DT