- Python 3.5+
- Keras 2.1.6 or higher
- pandas and pandas-ml
- TensorFlow 1.5 or higher
The dataset has 24 audio each for the sick and not sick people under data/* folder. https://osf.io/4pt2s/
In this example, only 2 classes will be picked for TensorFlow
not sick, sick
Data generation involves producing raw PCM wavform data containing a desired number of samples and at fixed sample rate and the following configuration is used
| Samples | Sample Rate | Clip Duration (ms) |
|---|---|---|
| 24 | 16000 | 5000 |
It is mostly a time stacked VGG-esque model with 1D Convolutions for temporal data such as audio waveforms. A one-dimensional dilated convolutional layer serving as the context convolution (context_conv) is employed to extract a wider field of the data. This is followed by a dimensionality reduction in the reduce_conv layer which employes 1-D MaxPooling to reduce the number of parameters passed to the subsequent layer.
The model can be trained by running train.py
python train.py [-h] [-sample_rate SAMPLE_RATE] \
[-batch_size BATCH_SIZE] \
[-output_representation OUTPUT_REPRESENTATION] \
-data_dirs DATA_DIRS [DATA_DIRS ...]
python train.py -sample_rate 16000 -batch_size 64 -output_representation raw -data_dirs data/train
After the training the model for 100 epochs, the following confusion matrix was generated for assessing classification performance.
| Predicted | silence | sick | not_sick |
|---|---|---|---|
| Actual | |||
| silence | 20 | 0 | 0 |
| sick | 0 | 16 | 8 |
| not_sick | 0 | 11 | 13 |
- Keras - Deep Learning Framework
- TensorFlow - Machine Learning Library
https://ai.googleblog.com/2017/08/launching-speech-commands-dataset.html