Skip to main content

Table 3 Comparison of EER for Mandarin and English frontends and different lattice reconstruction algorithms

From: Phone lattice reconstruction for embedded language recognition in LVCSR

Lattice reconstruction algorithm Mandarin EER (%) English EER (%)
A0. Word-based phone lattice 25.00 28.80
A1. Time-aligned phone lattice 21.69 22.75
A2. Phoneme-expanded 1-best lattice 21.43 21.71
A3. Frame-expanded N-best lattice 19.26 20.64
A3'. Pruned frame-expanded N-best lattice 18.94 20.35
PRVSM, standard PRVSM system 17.24 18.90
PRVSM-alt, PRVSM with LVCSR acoustic model 18.73 19.93
  1. EERs of the A3' system were obtained by using the 84% pruning beam width for Mandarin and 81% for English