-
Revert to v3 EOS head (retrained one fired immediately)
-
v7: EOS retrained on ground-truth frame counts from training data
-
Deploy v7: retrained EOS head (95% precision, 94% recall)
-
voices.json with conditioning_file for multi-turn
-
Fix EOS: restore v3 EOS head on cross-attn transformer
-
Deploy cross-attn 100K: P95 CER -40% vs v3
-
Revert to v3 checkpoint (ablation showed v4 was worse) + v3 voices
-
v3 voice KV caches (28 voices)
-
v4 voice: Michael_kv_layer3.npy
-
v4 voice: Michael_kv_layer2.npy
-
v4 voice: Michael_kv_layer1.npy
-
v4 voice: Michael_kv_layer0.npy
-
v4 voice: Maya_kv_layer5.npy
-
v4 voice: Maya_kv_layer4.npy
-
v4 voice: Maya_kv_layer3.npy
-
v4 voice: Maya_kv_layer2.npy
-
v4 voice: Maya_kv_layer1.npy
-
v4 voice: Maya_kv_layer0.npy
-
v4 voice: Lily_kv_layer5.npy
-
v4 voice: Lily_kv_layer4.npy
-
v4 voice: Lily_kv_layer3.npy
-
v4 voice: Lily_kv_layer2.npy
-
v4 voice: Lily_kv_layer1.npy
-
v4 voice: Lily_kv_layer0.npy
-
v4 voice: Liam_kv_layer5.npy
-
v4 voice: Liam_kv_layer4.npy
-
v4 voice: Liam_kv_layer3.npy
-
v4 voice: Liam_kv_layer2.npy
-
v4 voice: Liam_kv_layer1.npy
-
v4 voice: Liam_kv_layer0.npy
-
v4 voice: Lewis_kv_layer5.npy
-
v4 voice: Lewis_kv_layer4.npy
-
v4 voice: Lewis_kv_layer3.npy
-
v4 voice: Lewis_kv_layer2.npy
-
v4 voice: Lewis_kv_layer1.npy
-
v4 voice: Lewis_kv_layer0.npy
-
v4 voice: Jessica_kv_layer5.npy
-
v4 voice: Jessica_kv_layer4.npy
-
v4 voice: Jessica_kv_layer3.npy
-
v4 voice: Jessica_kv_layer2.npy
-
v4 voice: Jessica_kv_layer1.npy
-
v4 voice: Jessica_kv_layer0.npy
-
v4 voice: Isabella_kv_layer5.npy
-
v4 voice: Isabella_kv_layer4.npy
-
v4 voice: Isabella_kv_layer3.npy
-
v4 voice: Isabella_kv_layer2.npy
-
v4 voice: Isabella_kv_layer1.npy
-
v4 voice: Isabella_kv_layer0.npy
-
v4 voice: Heart_kv_layer5.npy
-
v4 voice: Heart_kv_layer4.npy
