Comparison with the other model

In addition, we compared the proposed model's performance with the performances of other typical methods for the above simulations. The other methods correspond to:

(1)

Previous model [Unoki and Akagi1999b],

(2)

Segregation model using constraints (ii-a) and (iii) (labeled by Cond. 1),

• extracting the harmonics using the Comb filter
• determining A_k(t) and $\theta_{1k}(t)$ from C_k,0(t) and D_k,0(t)

(3)

Segregation model using constraints (iii) (labeled by Cond. 2),

• extracting the harmonics using the Comb filter
• A_k(t)=S_k(t) and $\theta_{1k}(t)=\phi_k(t)$

(4)

Spectral subtraction [Boll1979] on the gammatone filterbank (labeled SS), and

• bias (mean of noise) subtraction
• half-wave rectification
• reduction of noise residual

(5)

Segregation model using no constraints (labeled by Cond. 3).

• all-pass filtering
• A_k(t)=S_k(t) and $\theta_{1k}(t)=\phi_k(t)$ for all k

We compared with the above conditions as follows:

(1)

can the proposed method improve our previous model,

(2)

does it have the advantage of the smoothness of the instantaneous amplitude and phase,

(3)

does it have the advantage of the instantaneous phase,

(4)

is it more useful than basic noise suppression, and

(5)

is the proposed method's segregation accuracy improved as noise separation.