Update README.md

README.md

@@ -17,7 +17,7 @@ Speech technology works well for a handful of languages and poorly for most of t
 
 We run public, human-rated evaluations of voice systems so that failures are visible and comparable, and we build curated datasets aimed at the failures those evaluations expose.
 
-The two feed each other. Evaluation tells us what models cannot do yet. That decides what data we collect next, and the data is released so anyone can train against the same hard cases.
+The two feed each other. Evaluation tells us what models cannot do yet. That decides what data we collect next, and the dataset is curated so anyone can train on it directly and see measurable gains in performace across benchmarks.
 
 ## Leaderboards