CU-MOVE: ADVANCED IN-VEHICLE SPEECH SYSTEMS FOR ROUTE NAVIGATION - DSP for In-Vehicle and Mobile Systems - page 39

where determines the degree of warping. For 16kHz sampled signals,

and

approximates the Mel and Bark scales, respectively.

4.3.3 PMVDR Algorithm

We can summarize the PMVDR algorithm as follows [37];

Step 1: Obtain the perceptually warped FFT power spectrum,
Step 2: Compute the “perceptual autocorrelations” by using IFFT on the
warped spectrum,
Step 3: Perform an Mth order LP analysis via Levinson-Durbin recursion
using perceptual autocorrelation lags [41],
Step 4: Calculate the Mth order MVDR spectrum using Eq. (7) from LP
coefficients [36],
Step 5: Obtain Cepstrum coefficients using the straightforward FFT-based
approach [43].

A flow diagram for the PMVDR algorithm is given in Figure 2-8. The
algorithm is integrated into the CU-Move recognizer as the default acoustic
feature front-end, (further information and code can be obtained from the CU-
Move web site [27]).

Figure 2-8. Flow Diagram of the PMVDR acoustic feature front-end

4.3.4 Experimental Evaluation

We evaluate the performance of PMVDR on the CU-Move extended digit
task [27,28,37] using our SONIC [23,25] LVCSR system. Sonic incorporates