espnet2.enh.layers.beamformer.get_rtf

source

espnet2.enh.layers.beamformer.get_rtf(psd_speech, psd_noise, mode='power', reference_vector: int | Tensor = 0, iterations: int = 3)

Calculate the relative transfer function (RTF).

The RTF is calculated using either the power method or eigenvalue decomposition. The algorithm for the power method is as follows:

1. rtf = reference_vector
2. for i in range(iterations): : rtf = (psd_noise^-1 @ psd_speech) @ rtf rtf = rtf / ||rtf||_2 # this normalization can be skipped
3. rtf = psd_noise @ rtf
4. rtf = rtf / rtf[…, ref_channel, :]

Note: Normalization at the reference channel is not performed here.

• Parameters:
  • psd_speech (torch.complex64/ComplexTensor) – Speech covariance matrix with shape (…, F, C, C).
  • psd_noise (torch.complex64/ComplexTensor) – Noise covariance matrix with shape (…, F, C, C).
  • mode (str) – One of (“power”, “evd”).
    • “power”: Uses the power method.
    • “evd”: Uses eigenvalue decomposition.
  • reference_vector (torch.Tensor or int) – Can be either a tensor of shape (…, C) or a scalar.
  • iterations (int) – Number of iterations to perform in the power method.
• Returns: The calculated RTF with shape (…, F, C, 1).
• Return type: rtf (torch.complex64/ComplexTensor)

Examples

>>> psd_s = torch.randn(10, 8, 4, 4, dtype=torch.complex64)
>>> psd_n = torch.randn(10, 8, 4, 4, dtype=torch.complex64)
>>> rtf = get_rtf(psd_s, psd_n, mode="power", iterations=5)
>>> print(rtf.shape)
torch.Size([10, 8, 4, 1])