MerlinPCarson / WakeWord-Detection

Training and evaluation scripts for wake word detection DNN models.
Apache License 2.0
8 stars 1 forks source link

summing wakeword posteriors, smoothing not wakeword posteriors, made … #18

Closed MerlinPCarson closed 3 years ago

MerlinPCarson commented 3 years ago

Changed how True Positives are counted, just sum over a threshold since there is now one posterior per wakeword utterance. Smoothed not wakeword posteriors. Removed results directory and save results to models directory. Changed from storing filtered frames from ring buffer to list, frames were being skipped with ring buffer. Removed overlapping from reading of samples. Made status bars in model evaluation nested. Created examine_audio flag for plotting misclassified wakewords with posteriors. Added estimate of wavenet results from Coucke et al. paper.