Digital Signal Processing Reference
In-Depth Information
for microphone channels
i
=2,3,4,5.
4.1.4
Experimental Evaluation
In order to evaluate the performance of the CSA-BF algorithms in noisy
car environments, we process all available speakers in Release 1.1a
[21,26,27] of the CU-Move corpus using both CSA-BF and DASB
algorithms, and compared the results. This release consists of 153 speakers, of
which 117 were from the Minneapolis, MN area. We selected 67 of these
speakers that include 28 males and 39 females, which reflects 8 hours of data.
In order to compare the result of CSA-BF with that of DASB thoroughly, we
also investigated the enhanced speech output from SA-BF. For evaluation, we
consider two different performance measures using CU-Move data. One
measure is the Segmental Signal-to-Noise Ratio (SegSNR) [22] which
represents a noise reduction criterion for voice communications. The second
performance measure is Word Error Rate (WER) reduction, which reflects
benefits for speech recognition applications. The Sonic Recognizer [23,25] is
used to investigate speech recognition performance. During the recognizer
evaluation, we used 49 speakers (23 male, 26 female) as the training set, and
18 speakers (13 male, 5 female) as the test set.
Table 2-1 summarizes average SegSNR improvement, average WER,
CORR (word correct rate), SUB (Word Substitution Rate), DEL (Word
Deletion Rate) and INS (Word Insertion Rate). Here, the task was on the
digits portion of CU-Move corpus (further details are presented in [19]).
Figure 2-5 illustrates average SegSNR improvement and WER speech
recognition performance results. The average SegSNR results are indicated by
the bars using the left-side vertical scale (dB), and the WER improvement is
the solid line using the right-side scale (%).
