Conclusion: Implications and Future Directions

Conclusion: Implications and Future Directions
The main conclusion is that the VLD works quite well, as long as it is evaluated sensibly.

Contributions of this thesis

improved method of evaluation using RVLMs
method for using MLP to combine knowledge based measurements

Enhancements to the VLD

peak detection can only get deletion rate down to 6%
current version has deletion rate of 8%
further improvement would require more sophisticated processing

System integration issues
Depending on how the VLD's output is used by the LAFF system,

optimization criteria may be adjusted
error correction by combining VLMs with CLMs
vowel feature analysis may be able to correct some insertions

System design issues

MLP to combine acoustic cues for distinctive features
acoustic cues shared between DFs

assume feature A to estimate feature B
assume feature B to estimate feature A
find combination with highest joint confidence