Machine Learning - Pitch Detection

• Input - continuous monophonic audio signal at 44.1kHZ at 16 bit
• ANN Architecture - Use an ANN specifically an RNN with LSTM or GRU trained using Adam.
• Layer (or should it simply be number of parameters) sized (ignore nyquist condition as included in sample rate) to capture longest wave form probably need a filter layer.
• Use RELU or a softmax output so there is a nice linear output.
• If each cell (4 weights) is a filter we could initialise forget weights to pick out frequencies rather than just being random.

$(\frac{1}{4}\cdot \frac{44100}{27.5})=400.9$

$(\frac{1}{4}\cdot 88)=22$

• Training Data - computer generated audio at various frequencies and phase shift possibly include harmonics and different wave forms. From A0 (27.5Hz) to B8 (7902.13Hz)
• Output - continuous value proportional to the frequency of the input signal
• Test data set - tuning fork, guitar strings, violin strings

Bibliography

1. A beginners Guide to LSTMs and Recurrent Neural Networks

Tweet