Digital Signal Processing Reference
In-Depth Information
modules. As expected, working partly in the frequency-domain and partly in
the time-domain the overall processing time has been
shortened
considerably.
Along the lines of many practical echo canceller systems reported in
literature, we have performed a frame-based updating strategy for our
Wiener filter coefficients as opposed to the sample-by-sample computations.
In addition to yielding considerable computational savings, this frame-by-
frame approach turns out to be consistent with the subsequent speech
compression stage, where one of the existing standards is normally employed
including LPC-10, RTP, CELP, MELP, and their derivatives.
EXPERIMENTS AND DISCUSSION OF RESULTS
4.
To evaluate the performance of the proposed technique the ASE-
MCMAC noise cancellation system of Figure 8-5, has been built as well as
the benchmark LMS-ASE system. The microphones and were
presented with a pre-recorded speech and noise data collected in a moving
vehicle subject to varying environmental conditions. We will refer these
microphones as the primary channel microphone which has been placed
towards the speaker and the noise reference microphone which was faced
away from the speaker, respectively. It is worth noting that this back-to-back
placement of microphones is the standard procedure in recoding studios.
Using several combinations of SNR values for and we have created
noisy speech, i.e., low-noise, medium-noise, and high-noise regimes.
In Figure 8-6, we illustrate the performance of our speech denoising
experiments based on the proposed MCMAC-ASE algorithm against the
benchmark LMS-ASE algorithm under various noise conditions.
