The speech corpus for the following validation example consists of un-supervised telephone recordings by 1000 speakers with the orthographic transcript as annotation.
Quantitative (Formal) Validation Procedure:
Check for 134 recording items per speaker.
Check for empty signal files.
Check for signals files with clippings; must always be less than 5%.
Check for S/N; must be more than 15 dB.
Check for correct terminology for all data files according to specs.
Check for one annotation file per signal file.
Check if annotation files are parsable.
Check for complete and parsable speaker profile per speaker.
Check for complete and parsable recording protocol per recording.
Check for 50/50% gender distribution +/- 5%.
Check for age distribution in two groups 18 - 32 and 32 - 64; both
groups have 50% +/- 5%.
Check for parsability and completeness of dictionary.
Qualitative Validation Procedure:
Check documentation for completeness and consistency.
Check 5% randomly selected annotation files by independent
manual transliteration and cross check results;
3% word errors (including insertions and
deletions) allowed.
Check 10% randomly selected entries from dictionary for correct
pronunciation; 2% phonemic errors (including insertions and deletions)
are allowed.