ROBUST ASR INSIDE A VEHICLE USING BLIND PROBABILISTIC BASED UNDER-DETERMINED CONVOLUTIVE MIXTURE SEPARATION TECHNIQUE - DSP for In-Vehicle and Mobile Systems - page 288

Figure 18-1. Scatter plot of data for BSS of three sources using 2 sensors. The circled x shows
the true value computed from the mixing matrix.

Fifty one utterances per speaker for each condition were collected. While
digitally recording this data it was sampled at 44.0 KHz. This data was then
down loaded to a computer in .wav format and was transcribed
orthographically. The speech data was downsampled to 8.0 KHz since the
segment based continuous speech recognizer that we used in our experiments
expects the data to be sampled at 8.0 KHz.

4. EXPERIMENTS

Speech recognition performance in terms of word recognition accuracy
percentage was obtained using the database both using the blind convolutive
mixture separation algorithm proposed above and without. An example of a
mixed speech signal from two channels and separated four speech signals
from the mixed signals using our approach is provided below in Figures 18-3