Figure 15.2 shows the results of the mean log-likelihood experiment. The vertical scale of the bar-graph equals $ \mu$$ _m$$ (DAF)$$ -$$ \mu$$ _m$$ (HMM)$. In the dyphthong data, the values are positive, indicating that DAF-HMM produces a higher log-likelihood. For office sounds data, DAF-HMM produces a lower log-likelihood. This indicates that the spectral content of the dyphthongs data was smoothly changing and feature time-derivatives are more meaningful from a statistical modeling point of view. For the abrupt sounds in the office sounds data, the feature time difference were random and not predictable. Thus, augmenting the features did more harm than good.
Figure: Difference in $ \mu _m$ for DAF-HMM-corrected and HMM for Dyphthongs (left) and Office sounds (right).

Below we list the results of the classification performance experiment.

Classification Error (percent)
Dyphthongs Office Sounds
9.6% 7.5% 0.66% 1.64%
The results are consistent with the mean log-likelihood experiment. They indicate that augmenting the feature is helpful for the Dyphthong data, but detrimental for the office sounds data.

Baggenstoss 2017-05-19