Digital Signal Processing Reference
In-Depth Information
Figure 18-4.
Separated four speech signals from two mixtures after applying our algorithm.
While applying the proposed convolutive mixture separation algorithm,
the speech data from two channels out of seven - one corresponding to
channel 1 of our microphone array and the second corresponding to Andrea
microphone array was considered. First, the speech data from channel 1 of
our microphone array was used to test the speech recognizer. Next, the
speech data that was enhanced by applying the mixture separation algorithm
was used. In the Table 18-1 word recognition accuracy for all the three
conditions - quiet, radio on and cross talk present with and without
enhancing speech signals is provided. From this table, it can be seen that
there is a significant improvement in speech recognition accuracy in all three
cases. In the case of quiet even though other sources such as radio, cross talk
was not present but due to the presence of car engine noise and road noise
the speech recognition accuracy degraded; however, when the mixture
separation algorithm was used to enhance the speech signals, the accuracy
improved significantly (15 %). Note that the number of utterances used for
each case is 102.
