Digital Signal Processing Reference
In-Depth Information
acoustic models. The overall critical performance rate (CPR from Eq. (6))
was calculated as 92.1%
Having established the environmental sniffer, and normalized cost matrix
for directing ASR model selection, we now turn to ASR system evaluation.
We tested and compared the following 3 system configurations: S1-model
matching was done using
a priori
knowledge of the acoustic noise condition
(i.e., establish theoretical best performance - matched noise conditions), S2-
model matching was done based on the environmental acoustic knowledge
extracted from Environmental Sniffing, S3-all acoustic condition dependent
models were used in a parallel multi-recognizer structure (e.g., ROVER)
without using any noise knowledge and the recognizer hypothesis with the
highest path score was selected.
Figure 2-7.
Word Error Rates for Digit Recognition Tests: S1 - matched noise model case, S2
- environmental sniffing model selection (1 CPU for sniffing, 1 CPU for ASR), S3 (ROVER) -
employs up to 9 recognizers (i.e., CPUs) trained for each noise condition with ROVER
selection.
