Digital Signal Processing Reference
In-Depth Information
Fig. 12.5 Microphone array aperture and location of signal sources for the RIR measurement
12.3 Experiments
To generate dual-channel speech signal and nonstationary interference signals,
room impulse responses (RIRs) were measured in a vehicular chamber which has
a reverberation time,
T
60
¼
250 ms. The desired speech source was modeled to be
located 50 cm from the microphone array along the broadside direction (90
) and
the nonstationary interference source to be 75 cm along the 45
line. The array was
located in front of the speech source with a 10-cm aperture. Figure
12.5
describes
the experimental setup for signal generation.
Each RIR was convoluted with a single-channel clean speech signal to produce a
dual-channel speech signal, and with an interfering human voice for a dual-channel
nonstationary interference noise at a sampling rate of 8 kHz. Brownian noise was
added as the in-vehicle ambient noise. Next, the interference plus the ambient noise
was combined with the speech signal to simulate various signals with interference
and noise ratios (SINR) ranging from
5 to 20 dB. The speech signal in
experiments was formed from Korean digit strings and a nonstationary interference
noise generated by using arbitrary Korean words.
To evaluate the performance of the noise suppression and the perceptibility of
the enhanced speech signal, the noise reduction (NR) in log-domain and the
perceptual evaluation of the speech (PESQ) are used as measures [
9
], respectively.
Table
12.1
shows the performance of the proposed dual-channel speech enhance-
ment system, where “PANS” and “PANS+ANC” denote the usage of PANS only
and PANS with the ANC to estimate the desired spectral envelopes, respectively.
The findings of the proposed algorithms is compared with the conventional transfer
function-based GSC (TFGSC) method [
2
].
As shown in Table
12.1
, the proposed PANS and PANS with ANC show superior
performance over that of the TFGSC. Although PANS shows similar speech quality
with TFGSC in adverse noise environment, this problem is solved by combining it
with an ANC.
Search WWH ::
Custom Search