A Likelihood-Maximizing Framework for Enhanced In-Car Speech Recognition Based on Speech Dialog System Interaction - Digital Signal Processing for In-Vehicle Systems and Safety - page 161

Table 10.2 ASR results for
increasing number of joint
optimization iterations

# Iterations IDL 35U 35D 55U 55D

Baseline 70.4 48.8 36.2 41.8 23.5

a ( k ) ¼ 1 73.3 47.8 36.8 44.5 26.1

1 74.1 49.1 38.1 45.1 26.4

2 74.1 49.4 37.7 44.8 26.1

3 73.9 49.9 37.2 44.8 26.0

4 74.0 50.1 37.2 44.5 26.3

5 74.0 50.3 37.1 44.4 26.1

10 74.1 50.2 37.5 44.1 25.9

one which is well below its best performance (0.8%) but still provides improvement
over the baseline and static MFNS systems. As a result, three gradient-descent
iterations have been used for the remainder of the experiments in this chapter.

10.4.2 Joint Optimization Iterations

Having established the most effective number of gradient-descent iterations, the
number of joint optimization iterations was analyzed. Table 10.2 shows these results
with the best performance across all noise conditions highlighted for clarity.
Apart from the 35-mph-with-windows-up noise condition, the results indicate
that only one joint optimization iteration is required for in-car speech recognition.
This result indicates that only minor changes are made to the decoded state
sequences and therefore appears to be no advantage in performing more than one
joint optimization iteration. Relating this observation to the results of the gradient-
descent iterations experiment, if the state sequence did not change at all, the
parameter optimization would continue from exactly the same position that it
finished previously, and therefore, over-optimization is likely to occur as the
number of joint optimization iterations increases.
This result combined with that of Sect. 10.4.1 indicates that over-optimization is a
serious issue for LIMA frameworks operating in vehicular environments. It is there-
fore suggested that optimization iterations be kept to a minimum in order to keep the
enhancement parameters generalized. The practical advantage of these findings is the
ability to achieve improved ASR using LIMA frameworks whilst creating minimal
processing delays due to the need for only a few optimization iterations.

10.4.3 LIMA Frameworks

The LIMA frameworks listed in Sect. 10.3.4 were tested using the results obtained
in the previous experiments. Table 10.3 presents the ASR results for all three