Digital Signal Processing Reference
In-Depth Information
d =
38 mm (a typical value for boomless headsets). Recordings were done in
the following situations:
Sitting in a shopping mall (Sit-Mall1)
Inside a working car parked next to a highway (HWY)
Inside a moving car with open windows (CarWOpen)
Inside a moving car with closed windows (CarWClose)
In each recording set, the first and second microphone inputs are
considered as primary and reference noises, respectively. Figures 10-5(a)-(d)
display the average Power Spectral Density (PSD) of the primary noise in
each case. While all recorded noise source spectra are lowpass, the PSDs of
the two Car noises (CarWOpen and CarWClose) fall much faster with
frequency than the others. The effect of engine noise mostly appears as two
local peaks at about 2.7 kHz and 5.4 kHz in Figures 10-5(b)-(d).
The spatial coherence curves of the first and second microphone signals
(representing signals
x
and
y
in Eq. (1)) in the above recording situations
are plotted in Figures 10-6(a)-(d), respectively. As depicted, the curves
closely match the theoretical spatial coherence computed for diffuse noise by
Eq. (2) with
d
= 38 mm. This observation is consistent with the diffuse noise
assumption usually considered for most environmental noise fields in
vehicular applications [2]. Also, due to its directivity, the engine noise acts
more as a coherent source evident from the peaks in the coherence function
at 2.7 kHz and 5.4 kHz (Figures 10-6(b)-(d)). This is especially evident for
noises in a moving car with closed windows (Figure 10-6(d)).
Several sentences from the TIMIT [14] database were used as the speech
material. For each set of the noise recordings, the primary noise was added to
speech signal at 0 dB SNR and utilized as the noisy input.
4.
HYBRID SPEECH ENHANCEMENT
EVALUATION
Through subjective and objective evaluations, the performance of the
SAFWF system is compared to that of the SAF approach. Also, we have
examined the performance of single-microphone standard Wiener filtering
(STDWF) method where the filter is directly applied to the primary noisy
input.
