Algorithm 1. Signal basis training using CNMF

Input: Spectrogram of signal dataset ${\mathit{Y}}_{\mathit{train}}$, priori basis number R, shift length of CNMF K, type of β-divergence β

number of iterations in CNMF $i_{\mathit{CNMF}}$

Output: The basis matrix of signal ${\mathit{H}}_{\tau }$

1:  Initialize ${\mathit{H}}_{\tau }$ and $\mathit{Q}$ with random non-negative values

2:  Normalize columns of ${\mathit{H}}_{\tau }$

3:  for i = 1, $\cdots$, $i_{\mathit{CNMF}}$ do

4:   for $\tau$= 0, $\cdots$, K − 1 do

5:    Compute Z

6:    Update ${\mathit{H}}_{\tau }$ and $\mathit{Q}$ using Equations (21) and (22)

7:    Normalize columns of ${\mathit{H}}_{\tau }$

8:   end for

9:  end for