To use the model, follow the steps below:
- Access the vneural.py, change the keyword at line 146, where collection of keyword label occurs Keyword given in example is 'nhu'
- Create folder 'train/' where contains keyword samples and non-keyword samples
- Run vneural.py and see how it works
Update 23-5-2019: Forget the steps above ...
- Have access to standardize.py for formatting .wav files to the standard form (i.e. to 44.1kHz wave and mix with silence recordings) Mean that you have to get at list 3 folder with: sliences, A_KEY_WORD, NON_KEY_WORDS folder, each contains samples representing the label
- Start training with train.py, just run train.py but mention the 'dictionairy' (folder:label) at line 98 which you have to configure them correctly. Once training completed, with automatical a file name 'model.keras' is created
- Use the 'model.keras' to test. Example testing is prepared in 'usage.py'. It both uses the trained neural network and Google speech recognition. Enjoy the experiments
Contact me at fb.me/vinhphuc.tadang if any issue exists or contribute to the project (not really a project :) )
When using prediction with sigmoid activation, use threshold 0.9 to judge the right word