Update README.md

README.md

@@ -172,7 +172,7 @@ that matter for separation difficulty — bilingual share and overlap share are
 identical across splits, so validation/test numbers are representative of the training
 distribution.
 
-![Split consistency](./figure/split_consistency.png)
+<img src="./figure/split_consistency.png" width="70%">
 
 ### 3. Controlled, full-range overlap
 
@@ -180,7 +180,7 @@ Overlap is not concentrated at one difficulty. Mixtures span the full range from
 clean (no overlap) to fully overlapped, with the harder end growing as speaker count
 increases (2→4 spk), which is the expected and desired behaviour.
 
-![Overlap distribution](./figure/overlap_distribution.png)
+<img src="./figure/overlap_distribution.png" width="70%">
 
 ### 4. Complete, machine-readable metadata
 
@@ -194,7 +194,7 @@ enabling exact, reproducible analysis of language / gender / overlap / timing.
 The share of bilingual (EN+KO) mixtures grows with speaker count, while EN-only /
 KO-only mixtures remain represented at every count:
 
-![Language composition](./figure/language_composition.png)
+<img src="./figure/language_composition.png" width="70%">
 
 ---