#48 Solution to TensorFlow Speech Recognition Challenge on small foot-print keyword spotting with neural networks.
To sort of reproduce my result:
- clone the repo
- download competition data and unzip to
inputdirectory . - run
run.sh.
I used a 1070 workstation and a couple of compute instances with K80 on Google Cloud Platform. For some model I used different machine to train different folds, collect them and run inference.py to get oof and test predictions.
Note: reproducibility is not 100% guaranteed, but you should be able to get models with 88% accuracy on public leaderboard and 89% accuracy on private leaderboard.
I have three different input features, MFCC, logspectrogram and logmelspectrogram.
For first layer models, I used the model archituctures described in Arms' paper Hello Edge: Keyword Spotting on Microcontrollers with a bit tweak of mine own. Also tried out 1d cnn and resnet.
List of first layer neural network archituctures:
- Fully connected network / MLP .
- 1D/2D VGG style Convolutional neural network .
- Recurrent neural network . (I wasn't using Cudnn RNN at the time, so limited experiments were done)
- Convolutional recurrent neural network .
- Depth seperatable convolutional neural network
- Joint model which concatenate intermediate layers from CNN and RNN .
- ResNet-18 and ResNet-34 .(My implementation is wrong)
My second layer stacker is a simple Ridge Classifier, and the stacking only improved very slightly over the best model.
Overall, it is sort of my first time tackling a deep learning competition, and I learned a lot during which.