The goal of this project is to build a voice recording digit classifier using the spoken digits dataset available here.

The classifier takes as an input any wave recording and will perform a speech/keyword detection algorithm outputting what digit is being spoken on the input recording.

The key part is using computer-vision techniques such as convolutional neural network to perform prediction, by leveraging the power and utility of mel spectrograms, as seen in the figure below, we are able to teach the machine "to see the sound recording" (an example of synesthesia).

Create a virtual environment (It is preferable as not avoid any dependecy problems) and install the required packages as mentioned in the requirements.txt file.

pip install -r requirements.txt

Once all the required packages are installed open the spoken_digits notebook and follow the the instructions to either train the model from scratch or load and test the trained model.

To prepare the dataset, first using tensorflow_datasets download the dataset and split as needed.

Take a moment to explore the dataset after reading more about it using the utils.py functions or by yourself.

The function preprocess.py will be particularely useful to convert all the recordings to a uniformly sized spectrogram images that will served as input for the CNN model.

Using the Keras library interface, we then build a sequential model with multiple CONV2D + MAXPOOling blocks.

Feel free to customize the layers and hyperparameters to see if you can get a higher accuracy!

When satisfied save yout model as done in the notebook. You will be able to load it again and use for prediction. You can also add a callback so that you can train it again.

Finally, you can test the model on your own recordings for fun.

As described in the last part of the notebook, upload your sample recordings into the directory and load them using librosa.load('path_to_filename', 'sample_rate') and specify the sample rate accordingly.

Then load your model, test and enjoy :) .