diff --git a/Android/src/app/src/main/AndroidManifest.xml b/Android/src/app/src/main/AndroidManifest.xml
index 5075c77ec..901cb4b8f 100644
--- a/Android/src/app/src/main/AndroidManifest.xml
+++ b/Android/src/app/src/main/AndroidManifest.xml
@@ -63,6 +63,9 @@
         android:supportsRtl="true"
         android:theme="@style/Theme.Gallery"
         tools:targetApi="31">
+        <uses-native-library android:name="libedgetpu_util.so" android:required="false" />
+        <uses-native-library android:name="libedgetpu_tachyon.google.so" android:required="false" />
+        <uses-native-library android:name="libedgetpu_client.google.so" android:required="false" />
         <!--
           android:configChanges="uiMode" tells the system don't destroy and
           recreate the activity when UI mode changes (e.g. setting dark mode).
diff --git a/Android/src/app/src/main/LLM_CAMERA.md b/Android/src/app/src/main/LLM_CAMERA.md
new file mode 100644
index 000000000..f6b5e6a6c
--- /dev/null
+++ b/Android/src/app/src/main/LLM_CAMERA.md
@@ -0,0 +1,14 @@
+# LlmCamera TPU Configuration
+
+To build the LlmCamera feature with TPU-accelerated VoiceLM TTS, you must pass the `voicelm_tpu` definition during your `blaze` build. This builds the **experimental** flavor of the AI Edge Gallery app (`ai_edge_gallery_app_experimental`), which is required because the LlmCamera feature is currently behind the experimental build target.
+
+**Example Build Command:**
+```bash
+blaze mobile-install --start_app -c opt --config=android_arm64-v8a --define=libunwind=true --define=xnnpack_use_latest_ops=true --define=keep_litertlm_symbols=true --define=voicelm_tpu=true //third_party/ai_edge_gallery/Android/src/app/src/main:ai_edge_gallery_app_experimental
+```
+
+**Important:** Before testing the TPU flavor on a device, you must enable unlisted apps to use the Edge TPU service. Run the following ADB command at least once on the test device:
+
+```bash
+adb shell setprop vendor.edgetpu.service.allow_unlisted_app true
+```
diff --git a/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/common/AudioAnimation.kt b/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/common/AudioAnimation.kt
index d99b4755d..3e6b8fd8d 100644
--- a/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/common/AudioAnimation.kt
+++ b/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/common/AudioAnimation.kt
@@ -19,6 +19,7 @@ package com.google.ai.edge.gallery.ui.common
 import android.graphics.RuntimeShader
 import android.os.Build
 import androidx.compose.animation.core.Animatable
+import androidx.compose.animation.core.LinearOutSlowInEasing
 import androidx.compose.animation.core.tween
 import androidx.compose.foundation.Canvas
 import androidx.compose.foundation.layout.fillMaxSize
@@ -36,6 +37,15 @@ import androidx.compose.ui.graphics.ShaderBrush
 import kotlin.math.pow
 import kotlin.random.Random
 
+private const val AMPLITUDE_VISUAL_SCALE_DOWN = 0.4f
+// Maximum amplitude value from LlmCameraViewModel.convertRmsDbToAmplitude (0..65535 range).
+// Used to normalize the amplitude input to 0..1 for the visual animation.
+private const val MAX_16BIT_AMPLITUDE = 65535.0
+// Normalized amplitude threshold for detecting high-to-low transitions that reset Perlin noise.
+private const val AMPLITUDE_THRESHOLD = 0.2
+// Multiplier for the random Perlin noise offset range.
+private const val POFFSET_RANGE = 1000f
+
 private const val SHADER =
   """
 // The size of the render area.
@@ -89,20 +99,23 @@ half4 main(float2 fragCoord) {
 
   // Add a wavy distortion to the y-coordinate of the uv.
   //
-  // Control the amplitude of the wave
+  // Control the amplitude of the wave. Higher values create more pronounced vertical distortion.
   float wave_strength = 0.036;
-  // Control the speed of the wave
-  float wave_speed = 1.2;
-  // Control the frequency of the wave
-  float wave_frequency = 4.0;
+  // Control the speed of the wave. Higher values make the idle wave animate faster.
+  float wave_speed = 0.8;
+  // Control the frequency of the wave. Higher values create more wave cycles across the screen.
+  float wave_frequency = 0.5;
 
   // Idle.
   if (amplitude == 0.) {
     uv.y += sin(uv.x * wave_frequency + -iTime * wave_speed) * wave_strength;
   }
   // Visualizing amplitude by sampling the 1d perlin noise at the given offset.
+  // The 0.5 multiplier on uv.x reduces the noise sampling rate, creating wider, smoother
+  // wave patterns. Dividing amplitude by 2.0 scales the visual response to prevent
+  // excessive distortion at high audio amplitudes.
   else {
-    uv.y -= perlin_noise_1d(pOffset + uv.x * 3.) * amplitude / 2.0;
+    uv.y -= perlin_noise_1d(pOffset + uv.x * 0.5) * amplitude / 2.0;
   }
 
   vec3 col = mix4(
@@ -142,8 +155,11 @@ fun AudioAnimation(bgColor: Color, amplitude: Int, modifier: Modifier = Modifier
     var iTime by remember { mutableFloatStateOf(0f) }
     var curPOffset by remember { mutableFloatStateOf(0f) }
     var prevNormalizedAmplitude by remember { mutableDoubleStateOf(0.0) }
-    // Use pow(x, 0.5) to make low amplitude levels more significant.
-    val normalizedAmplitude = (amplitude / 32767.0).pow(0.5)
+    // Use pow(x, 0.5) (square root curve) to make low amplitude levels more significant.
+    // Scale down by AMPLITUDE_VISUAL_SCALE_DOWN (0.4) to reduce the visual intensity,
+    // preventing the animation from being too aggressive at normal speech volumes.
+    val normalizedAmplitude =
+      (amplitude / MAX_16BIT_AMPLITUDE).pow(0.5) * AMPLITUDE_VISUAL_SCALE_DOWN
     var animatedAmplitude by remember { mutableFloatStateOf(normalizedAmplitude.toFloat()) }
 
     // Animate the amplitude value whenever amplitude changes.
@@ -152,7 +168,7 @@ fun AudioAnimation(bgColor: Color, amplitude: Int, modifier: Modifier = Modifier
       val animatable = Animatable(initialValue = animatedAmplitude)
       animatable.animateTo(
         targetValue = normalizedAmplitude.toFloat(),
-        animationSpec = tween(durationMillis = 100),
+        animationSpec = tween(durationMillis = 500, easing = LinearOutSlowInEasing),
       ) {
         animatedAmplitude = this.value
       }
@@ -160,8 +176,10 @@ fun AudioAnimation(bgColor: Color, amplitude: Int, modifier: Modifier = Modifier
 
     // Updates the iTime uniform for the shader.
     LaunchedEffect(Unit) {
-      while (true) {
-        withFrameMillis { frameTimeMs -> iTime = frameTimeMs / 1000f }
+      if ("robolectric" != Build.FINGERPRINT) {
+        while (true) {
+          withFrameMillis { frameTimeMs -> iTime = frameTimeMs / 1000f }
+        }
       }
     }
 
@@ -171,8 +189,10 @@ fun AudioAnimation(bgColor: Color, amplitude: Int, modifier: Modifier = Modifier
       // level (0.2 or greater) to a low level (less than 0.2). This makes the noise-driven visual
       // effect appear to "jump" or reset to a new, random state when the audio becomes quiet,
       // preventing the visual from settling into a repetitive or static pattern.
-      if (normalizedAmplitude < 0.2 && prevNormalizedAmplitude >= 0.2) {
-        curPOffset = Random.nextFloat() * 1000f
+      if (
+        normalizedAmplitude < AMPLITUDE_THRESHOLD && prevNormalizedAmplitude >= AMPLITUDE_THRESHOLD
+      ) {
+        curPOffset = Random.nextFloat() * POFFSET_RANGE
       }
       prevNormalizedAmplitude = normalizedAmplitude
 
diff --git a/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatModelHelper.kt b/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatModelHelper.kt
index 838cc28df..cee4c43a1 100644
--- a/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatModelHelper.kt
+++ b/Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatModelHelper.kt
@@ -273,7 +273,11 @@ object LlmChatModelHelper : LlmModelHelper {
 
   override fun stopResponse(model: Model) {
     val instance = model.instance as? LlmModelInstance ?: return
-    instance.conversation.cancelProcess()
+    try {
+      instance.conversation.cancelProcess()
+    } catch (e: IllegalStateException) {
+      Log.w(TAG, "Conversation is not alive, cannot cancel process", e)
+    }
   }
 
   override fun runInference(
diff --git a/Android/src/app/src/main/res/values/strings.xml b/Android/src/app/src/main/res/values/strings.xml
index 87d61da42..e52612417 100644
--- a/Android/src/app/src/main/res/values/strings.xml
+++ b/Android/src/app/src/main/res/values/strings.xml
@@ -125,9 +125,11 @@
   <string name="listening" translatable="true">Listening...</string>
   <string name="llm_camera_status_waiting_to_start" translatable="true" description="Status text shown when LlmCamera is initialized and waiting for user input.">Waiting to start…</string>
   <string name="llm_camera_status_loading_model" translatable="true" description="Status text shown while loading the LlmCamera model.">Loading model…</string>
+  <string name="llm_camera_status_listening" translatable="true" description="Status text shown when LlmCamera is listening to user speech. [CHAR_LIMIT=30] [BACKUP_MESSAGE_ID: 4062899179021837624]">Listening…</string>
   <string name="llm_camera_status_processing" translatable="true" description="Status text shown while LlmCamera is processing speech or image input.">Processing…</string>
   <string name="llm_camera_status_cancelled" translatable="true" description="Status text shown when LlmCamera processing was cancelled.">Cancelled</string>
   <string name="llm_camera_status_responding" translatable="true" description="Status text shown while LlmCamera model is generating a response.">Responding…</string>
+  <string name="llm_camera_status_done" translatable="true" description="Status text shown when LlmCamera finishes generating a response. [CHAR_LIMIT=15] [BACKUP_MESSAGE_ID: 5969525977325895045]">Done</string>
   <string name="llm_camera_status_error" translatable="true" description="Status text template shown when an error occurs during LlmCamera operation.">Error: %1$s</string>
   <string name="llm_camera_recognized_prefix" translatable="true" description="Prefix for recognized speech text in LlmCamera.">Recognized: %1$s</string>
   <string name="llm_camera_response_prefix" translatable="true" description="Prefix for model response text in LlmCamera.">Response: %1$s</string>
@@ -641,4 +643,7 @@
   <string name="config_label_font_size" translatable="true">Font size</string>
   <!-- Label for the 'Max character count' option in the model configuration dialog. [CHAR_LIMIT=NONE] -->
   <string name="config_label_max_char_count" translatable="true">Max character count</string>
+
+  <string name="llm_camera_title" translatable="true" description="Title of the LlmCamera screen. [CHAR_LIMIT=30]">LlmCamera</string>
+  <string name="llm_camera_cd_flip_camera" translatable="true" description="Content description for the flip button in LlmCamera. Camera is a generic term. [CHAR_LIMIT=30]">Flip Camera</string>
 </resources>