By deliberately varying some aspects of the system, you can discover how much effect they have on the overall quality of the voice.
A nice technique that you can use to discover the importance (or otherwise) of some aspects of the system is to deliberately introduce errors. Here are some ideas to get you started:
- produce less accurate pitch marks by varying the parameters to the pitch marker
- investigate the effect of alignment quality independently from database size: train the HMMs used for alignment on only a subset of the data, then use them to align the whole database (requires modifications to the
do_alignmentscript) - investigate the effectiveness of the vowel reduction stage in forced alignment (requires modifications to the
phone_substitutionsfile in thealignmentfolder) - use the ‘wrong’ dictionary