Hi, I have been trying to denoise a speech signal using Kalman Filter. I have use the following code for my calculation. May I know if I am in the right direction? Many literature that I have read uses a state space model for the Kalman Filter? How can I do that for a corrupted speech signal? System Identification?

Thanks. 

def my_calculation(y_corrupt):

  sum1 = 0

  square1 = 0

  mean1 = 0

  E_estimated = 0.00001

  E_measured = 0.00001

  noise_model= np.zeros((len(y_corrupt)))

  y_filtered = np.zeros((len(y_corrupt)))

  N = len(y_corrupt)

  for i in range (2,N):

    KG = E_estimated/(E_estimated+E_measured)

    #KG = KG**2

    noise_model[i] = noise_model[i-1] + KG*(y_corrupt[i]-noise_model[i-1])

    E_estimated= (1-KG)*E_measured;

    #E_estimated= (1-KG)*E_estimated;

    square1 = square1 + noise_model[i]**2

    sum1 = sum1 + noise_model[i]

    mean1 = (mean1*(i-1)+noise_model[i])/i

    E_measured = square1/i - mean1**2

    y_filtered[i] = y_corrupt[i] - noise_model[i]

  return y_filtered

I suggest you see the notebook https://github.com/strongio/torch-kalman for a general application. You can customize it for your data.

Hi,

You may check these links;

https://arxiv.org/ftp/arxiv/papers/1204/1204.0375....

https://github.com/montanerk/kalman-denoising

https://github.com/Vishnu-Dhanabalan/Speech-Enhanc...