Sync model card with upstream GitHub inference README

README.md

@@ -116,14 +116,15 @@ friendly, but quantized weights have not yet been calibrated and released.
 
 ### Synthetic validation split
 
-1000 clips drawn from DNS5 + AEC Challenge synthetic data, AECMOS over
-a 100-clip sub-sample per the standard AEC Challenge protocol.
+500 clips drawn from DNS5 + AEC Challenge synthetic data, stratified
+across 15 scenario cells. AECMOS over a 100-clip sub-sample per the
+standard AEC Challenge protocol.
 
 | Metric | Value |
 |---|---:|
-| ERLE (dB) | 15.44 |
+| ERLE (dB) | 11.4 |
 | AECMOS echo (↑, 1–5) | 3.83 |
-| AECMOS degradation (↑, 1–5) | 3.91 |
+| AECMOS degradation (↑, 1–5) | 4.04 |
 
 - **ERLE** (Echo Return Loss Enhancement) — `10·log10(E[mic²] / E[enh²])`
   averaged across scenarios. On scenes with active near-end speech both
@@ -140,11 +141,11 @@ Stratified 150-sample eval (30 per scenario) on the
 
 | Scenario | AECMOS echo | AECMOS deg | blind ERLE |
 |---|---:|---:|---:|
-| doubletalk | 4.82 | 2.54 | 12.5 dB |
-| doubletalk-with-movement | 4.76 | 2.47 | 9.7 dB |
-| farend-singletalk | 3.45 | 4.93 | 47.7 dB |
-| farend-singletalk-with-movement | 3.96 | 4.93 | 52.5 dB |
-| nearend-singletalk | 4.94 | 3.83 | 4.7 dB |
+| doubletalk | 4.73 | 2.09 | 11.4 dB |
+| doubletalk-with-movement | 4.72 | 2.35 | 8.3 dB |
+| farend-singletalk | 4.03 | 4.95 | 42.9 dB |
+| farend-singletalk-with-movement | 4.26 | 4.82 | 48.2 dB |
+| nearend-singletalk | 4.95 | 3.98 | 4.2 dB |
 
 ### Why DNSMOS OVRL is not reported here