diff --git a/packages/stage-ui-three/src/components/Model/VRMModel.vue b/packages/stage-ui-three/src/components/Model/VRMModel.vue
index b721c43de..57fc6d0ed 100644
--- a/packages/stage-ui-three/src/components/Model/VRMModel.vue
+++ b/packages/stage-ui-three/src/components/Model/VRMModel.vue
@@ -422,7 +422,7 @@ async function loadModel() {
         blink.update(vrm.value, delta)
         idleEyeSaccades.update(vrm.value, lookAtTarget, delta)
         vrmEmote.value?.update(delta)
-        vrmLipSync.update(vrm.value)
+        vrmLipSync.update(vrm.value, delta)
         vrm.value?.springBoneManager?.update(delta)
       }).off
 
diff --git a/packages/stage-ui-three/src/composables/vrm/lip-sync.ts b/packages/stage-ui-three/src/composables/vrm/lip-sync.ts
index cac786d83..ae8a3f33f 100644
--- a/packages/stage-ui-three/src/composables/vrm/lip-sync.ts
+++ b/packages/stage-ui-three/src/composables/vrm/lip-sync.ts
@@ -15,30 +15,111 @@ export function useVRMLipSync(audioNode: Ref<AudioBufferSourceNode | undefined,
   const { state: lipSyncNode, isReady } = useAsyncState(createWLipSyncNode(audioContext, profile as Profile), undefined)
 
   // https://github.com/mrxz/wLipSync/blob/c3bc4b321dc7e1ca333d75f7aa1e9e746cbbb23a/example/index.js#L50-L66
-  const lipSyncMap = {
+  const RAW_KEYS = ['A', 'E', 'I', 'O', 'U', 'S'] as const
+  type LipKey = 'A' | 'E' | 'I' | 'O' | 'U'
+  const LIP_KEYS: LipKey[] = ['A', 'E', 'I', 'O', 'U']
+  const BLENDSHAPE_MAP: Record<LipKey, string> = {
     A: 'aa',
     E: 'ee',
     I: 'ih',
     O: 'oh',
     U: 'ou',
   }
+  const RAW_TO_LIP: Record<typeof RAW_KEYS[number], LipKey> = {
+    A: 'A',
+    E: 'E',
+    I: 'I',
+    O: 'O',
+    U: 'U',
+    S: 'I',
+  }
+
+  const smoothState: Record<LipKey, number> = { A: 0, E: 0, I: 0, O: 0, U: 0 }
+  const ATTACK = 50 // the speed moving to the next mouth shape animation
+  const RELEASE = 30 // the speed ending the current mouth shape animation
+  const CAP = 0.7
+  const SILENCE_VOL = 0.04
+  const SILENCE_GAIN = 0.05
+  const IDLE_MS = 160
+  let lastActiveAt = 0
 
   watch([isReady, audioNode], ([ready, newAudioNode], [, oldAudioNode]) => {
-    if (oldAudioNode)
-      oldAudioNode.disconnect()
-
-    if (ready && newAudioNode)
-      newAudioNode.connect(lipSyncNode.value!)
+    if (oldAudioNode && oldAudioNode !== newAudioNode) {
+      try {
+        oldAudioNode.disconnect()
+      }
+      catch {}
+    }
+    if (!ready || !newAudioNode || !lipSyncNode.value)
+      return
+    try {
+      newAudioNode.connect(lipSyncNode.value)
+    }
+    catch {}
   }, { immediate: true })
   onUnmounted(() => audioNode.value?.disconnect())
 
-  function update(vrm?: VRMCore) {
-    if (!vrm?.expressionManager || !lipSyncNode.value)
+  function update(vrm?: VRMCore, delta = 0.016) {
+    const node = lipSyncNode.value
+    if (!vrm?.expressionManager || !node)
       return
 
-    for (const key of Object.keys(lipSyncNode.value.weights)) {
-      const weight = lipSyncNode.value.weights[key] * lipSyncNode.value.volume
-      vrm.expressionManager?.setValue(lipSyncMap[key as keyof typeof lipSyncMap], weight)
+    const vol = node.volume ?? 0
+    const amp = Math.min(vol * 0.9, 1) ** 0.7
+
+    // Remapping wLipSync output AEIOUS to AEIOU
+    const projected: Record<LipKey, number> = { A: 0, E: 0, I: 0, O: 0, U: 0 }
+    for (const raw of RAW_KEYS) {
+      const lip = RAW_TO_LIP[raw]
+      const rawVal = node.weights[raw] ?? 0
+      projected[lip] = Math.max(projected[lip], rawVal * amp)
+    }
+
+    // winner + runner
+    // Original code: all AEIOU mouth shape blended together. Because the A mouth shape has the largest deformation, mixing A-E-I-O-U based on their raw weights causes the combined result to be biased heavily toward A in most cases.
+    // Improved code: Only the 2 mouth shapes with the largest weights will be blended.
+    let winner: LipKey = 'I'
+    let runner: LipKey = 'E'
+    let winnerVal = -Infinity
+    let runnerVal = -Infinity
+    for (const key of LIP_KEYS) {
+      const val = projected[key]
+      if (val > winnerVal) {
+        runnerVal = winnerVal
+        runner = winner
+        winnerVal = val
+        winner = key
+      }
+      else if (val > runnerVal) {
+        runnerVal = val
+        runner = key
+      }
+    }
+
+    // Detect pause or keep silence
+    const now = performance.now()
+    let silent = amp < SILENCE_VOL || winnerVal < SILENCE_GAIN
+    if (!silent)
+      lastActiveAt = now
+    if (now - lastActiveAt > IDLE_MS)
+      silent = true
+
+    // winner + runner weights
+    const target: Record<LipKey, number> = { A: 0, E: 0, I: 0, O: 0, U: 0 }
+    if (!silent) {
+      target[winner] = Math.min(CAP, winnerVal)
+      target[runner] = Math.min(CAP * 0.5, runnerVal * 0.6)
+    }
+
+    // smoothness and expression generation
+    for (const key of LIP_KEYS) {
+      const from = smoothState[key]
+      const to = target[key]
+      // lerp
+      const rate = 1 - Math.exp(-(to > from ? ATTACK : RELEASE) * delta)
+      smoothState[key] = from + (to - from) * rate
+      const weight = (smoothState[key] <= 0.01 ? 0 : smoothState[key]) * 0.7
+      vrm.expressionManager.setValue(BLENDSHAPE_MAP[key], weight)
     }
   }
 
diff --git a/packages/stage-ui/src/stores/providers/openai-compatible-builder.ts b/packages/stage-ui/src/stores/providers/openai-compatible-builder.ts
index 2959e864f..faaf743fe 100644
--- a/packages/stage-ui/src/stores/providers/openai-compatible-builder.ts
+++ b/packages/stage-ui/src/stores/providers/openai-compatible-builder.ts
@@ -289,12 +289,12 @@ export function buildOpenAICompatibleProvider(
     validators: finalValidators,
     ...(resolvedCategory === 'transcription'
       ? {
-        transcriptionFeatures: transcriptionFeatures ?? {
-          supportsGenerate: true,
-          supportsStreamOutput: false,
-          supportsStreamInput: false,
-        },
-      }
+          transcriptionFeatures: transcriptionFeatures ?? {
+            supportsGenerate: true,
+            supportsStreamOutput: false,
+            supportsStreamInput: false,
+          },
+        }
       : {}),
     ...rest,
   } as ProviderMetadata