Classifying Mandarin tones using machine learning
Two Jupyter notebooks detailing an attempt to classify the tones of a Mandarin Chinese utterance based on the audio spectrogram.
Part 1 uses a manual extraction of fetures from extracted tone lines; Part 2 applies the full spectrogram as the input of a
convolutional neural network; Part 3 uses a recurrent neural network to encode the variable length input spectrogram.
The model obtained in Part 3 is used for tone clasificaiton in the related demo Android application.