Kaggle TensorFlow Speech Recognition Challenge
This repository contains my solution to the Kaggle TensorFlow Speech Recognition Challenge. The competition's goal was to train a model to recognize ten simple spoken words using Google Brain's speech command data set.
I wrote my solution in Python and TensorFlow. As a learning exercise, I tried designing my own network rather than starting with one of the well known architectures. I started with a pure RNN approach but later added 2D convolutional layers as they were effective in finding structure within the Mel Frequency Cepstral Coefficients transformation of the audio signals. My final design was:
| Layer | Type | Description |
|---|---|---|
| Input | Data | MFCC transformed data |
| Conv1 | Conv2d | 64 filters, 10x5 kernel, 1x2 strides |
| BatchNorm1 | Batch Norm | Momentum 0.9 |
| Relu1 | Relu | Relu activation layer |
| MaxPool1 | MaxPool | 3x3 kernel, 1x2 strides |
| DropOut1 | Dropout | Rate 0.5 |
| Conv2 | Conv2d | 128 filters, 10x5 kernel, 1x2 strides |
| BatchNorm2 | Batch Norm | Momentum 0.9 |
| Relu2 | Relu | Relu activation layer |
| MaxPool2 | MaxPool | 3x3 kernel, 1x2 strides |
| DropOut2 | Dropout | Rate 0.5 |
| Conv3 | Conv2d | 256 filters, 10x5 kernel, 1x1 strides |
| BatchNorm3 | Batch Norm | Momentum 0.9 |
| Relu3 | Relu | Relu activation layer |
| MaxPool3 | MaxPool | 3x5 kernel, 1x1 strides |
| DropOut3 | Dropout | Rate 0.5 |
| RNN1 | GRU | 128 units |
| FC1 | Dense | 256 units |
| BatchNorm 4 | Batch Norm | Momentum 0.9 |
| Relu4 | Relu | Relu activation layer |
| FC2 | Dense | Final output logits |
I trained the model for 30 epochs at a batch size of 512 on my NVIDIA TitanX GPU. I used the Python Speech Features package to transform the data from audio to MFCC. My final score was 80.8% accuracy which placed me 479th out of 1315 teams (top 37%).