Accuracy on testing_list.txt

ARM-software / ML-KWS-for-MCU

Keyword spotting on Arm Cortex-M Microcontrollers

Apache License 2.0

1.14k stars 417 forks source link

Accuracy on testing_list.txt #61

Closed LeighSu closed 5 years ago

LeighSu commented 6 years ago

I have run the predefined model on the files named in testing_list.txt(under the speech_dataset), and the accuracy are 75.9% 79.2% 81.0% for DNN_S, DNN_M, DNN_L model, 86.3% 87.6% 90.4% for CNN_S, CNN_M, CNN_L model.

navsuda commented 6 years ago

@LeighSu, I see the difference is because of (a) it was trained and tested on speech_commands_v0.01 (i.e. previous release of the dataset) and (b) the separation of dataset to train/val/test is based on a hash of the audio file names (see Custom Training Data section in this tensorflow tutorial), which may not exactly match with testing_list.txt.