3 ms·

Or you could let the neural network solve the inverse problem for you, at least for speech generation.