Low Correlation in Completeness Calculation

[teinhonglo]

Hello,

I am currently working on calculating completeness as shown in the documentation (score range: 0.0 – 1.0, representing the percentage of words with good pronunciation). My current approach is as follows:

From score.json, I calculate word-level accuracy using the ratio of the number of words that score greater than 7, 8, or 9 points as a threshold for "good pronunciation." Then, I compute the Pearson Correlation Coefficient (PCC) between the calculated word-level accuracy and completeness. Unfortunately, I am getting a very low correlation (around 0.2), which makes me question if my validation approach is correct.

My questions are:

Is my validation method for calculating word-level accuracy against completeness appropriate? Is there a more natural or standard way to calculate completeness that might yield better results? Any feedback or guidance on this would be highly appreciated!

Thank you!