Skip to main content

Table 3 Effect of the various choices in the design of the system’s first stage (discussed in Section 4.4) to the room-localized SAD performance on the DIRHA-sim test set

From: Room-localized speech activity detection in multi-microphone smart homes

Oper

\({\mathcal {M}}\)

Classes \({\mathcal {J}}\)

Recall

Precision

F-score

RI

\({\mathcal {M}}_{\mathrm {\,all}}\!\!\!\!\)

{ spall,silall}

72.30

56.63

63.51

RL

\({\mathcal {M}}_{\,r}\)

{ spr,silall}

72.07

61.08

66.12

  

{ spr,silr}

71.20

60.39

65.35

  

\(\{\,{\text {sp}}_{\,r\,},{\text {sp}}_{\,{\bar {r}}\,},{\text {sil}}_{\,\text {all}\,}\}\)

71.00

62.40

66.43

  1. For consistency, the first stage is always followed by the second stage of the MFCC/GMM baseline of Section 6.1. RI denotes room-independent operation (“oper”) of the first stage and RL room-localized one