1. WPE

class torchiva.WPE(n_iter=3, n_delay=3, n_taps=5, model=None, eps=1e-05)

Weighted prediction error (WPE) 9.

Parameters

• n_iter (int, optional) – The number of iterations. (default: 3)

• n_taps (int, optional) – The length of the dereverberation filter (default: 5).

• n_delay (int, optional) – The number of delay for dereverberation (default: 3).

• model (torch.nn.Module, optional) – The model of source distribution. If None, time-varying Gaussian is used. (default: None).

• eps (float, optional) – A small constant to make divisions and the like numerically stable (default:1e-5).

Returns

Y – The dereverberated signal in STFT-domain.

Return type

torch.Tensor, shape (..., n_src, n_freq, n_frames)

References

9

T. Nakatani, T. Yoshioka, K. Kinoshita, M. Miyoshi, and B. H. Juang, “Speech dereverberation based on variance-normalized delayed linear prediction”, IEEE Trans. on Audio, Speech, and Lang. Process., 2010.