Table 2 |
||
|
Experimental configurations |
||
|
Backend ID |
Frontend type |
Lattice reconstruction algorithm for VSM training and test |
|
|
||
|
A0 |
LVCSR |
None (using word-based phone lattices) |
|
A1 |
LVCSR |
Time-aligned phone lattice |
|
A2 |
LVCSR |
Phoneme-expanded 1-best lattice |
|
A3 |
LVCSR |
Frame-expanded N-best lattice |
|
A3' |
LVCSR |
Pruned frame-expanded N-best lattice |
|
PRVSM |
Phone-loop |
None (standard PRVSM system) |
|
PRVSM-alt |
Phone-loop |
None (standard PRVSM system with LVCSR acoustic model) |
|
|
||
|
Each experiment was implemented twice, once with an English LVCSR frontend and once with a Mandarin LVCSR frontend |
||
|
Shan et al. EURASIP Journal on Audio, Speech, and Music Processing 2012 2012:15 doi:10.1186/1687-4722-2012-15 |
||
