Unofficial implementation of Broadcasted Residual Learning for Efficient Keyword Spotting
- add specaug to train
Train
; train scaled 2 times model for 50 epochs and save best checkpoint to model-sc-2.pt
python main.py train --scale 2 --epoch 50 --checkpoint-file model-sc-2.pt
; Device: cuda
; Use subspectral norm: True
; --- start epoch 0 ---
; Train Epoch: 0 Loss: 3.6272
; Train Epoch: 0 Loss: 1.6613
; ...
; Train Epoch: 49 Loss: 0.3026
; Validation accuracy: 0.9626289950906722
; Top validation accuracy: 0.9628293758140467
; Test accuracy: 0.9604725124943208
Test
; test saved model on test dataset
python main.py test --scale 2 --model-file model-sc-2.pt
; Test accuracy: 0.9604725124943208
Apply
; apply saved model to wav file
python main.py apply --scale 2 --model-file model-sc-2.pt --wav-file SpeechCommands/speech_commands_v0.02/seven/5744b6a7_nohash_0.wav
seven 0.99977
six 0.00011
stop 0.00008
happy 0.00002
up 0.00000
You can find pretrained model-sc-2.pt model in example_model folder.
Use help
python main.py --help
python main.py train --help
python main.py test --help
python main.py apply --help
This implementation use all 35 labels from Google Speech Commands Dataset. Original paper use 10 commands and additional re-balanced "Unknown word" and "Silence" labels (section 4.1 in paper).