Now we compare MFCC feature chains with an AR feature chain
using data from both AR and MFCC-based circularly-stationary models.
In Figure 11.3, we show the acid test results for
data generated according to the circular AR model.
The accuracy of the AR feature chain is far superior.
Figure 11.3:
Acid test performance of AR features (left)
and MFCC features (right) for data generated according to the
circular AR model.
|
In Figure 11.4, we show the acid test results for
data generated according to the circular MFCC model.
The accuracy of the MFCC feature chain is far superior.
Figure 11.4:
Acid test performance of AR features (left)
and MFCC features (right) for data generated according to the
circular AR model.
|
These experiments underscore the importantce of
matching the feature extraction to the data.
Notice also that by using the “wrong"
freatures, the result is PDF estimation
accuracy. For the mis-matched feature,
log-likelihood for the
projected PDF is always biased significantly lower, never
higher! This is because the true PDF is the likelihood
function with the highest average likelihood.