diff --git a/app/src/androidTest/java/com/opendash/app/e2e/FakeSttPipelineE2ETest.kt b/app/src/androidTest/java/com/opendash/app/e2e/FakeSttPipelineE2ETest.kt
new file mode 100644
index 00000000..61971d66
--- /dev/null
+++ b/app/src/androidTest/java/com/opendash/app/e2e/FakeSttPipelineE2ETest.kt
@@ -0,0 +1,92 @@
+package com.opendash.app.e2e
+
+import androidx.test.ext.junit.runners.AndroidJUnit4
+import com.google.common.truth.Truth.assertThat
+import com.opendash.app.e2e.fakes.FakeSpeechToText
+import com.opendash.app.voice.stt.SpeechToText
+import com.opendash.app.voice.stt.SttResult
+import dagger.hilt.android.testing.HiltAndroidRule
+import dagger.hilt.android.testing.HiltAndroidTest
+import kotlinx.coroutines.flow.toList
+import kotlinx.coroutines.test.runTest
+import org.junit.Before
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import javax.inject.Inject
+
+/**
+ * Verifies the [SpeechToText] @TestInstallIn swap works end-to-end:
+ * the Hilt graph hands every consumer the [FakeSpeechToText] singleton,
+ * and queued results flow through to a collector.
+ *
+ * `VoicePipeline.startListening()` itself is service-aware (audio focus,
+ * VoiceService wake-word pause, beep playback) so wiring the **full**
+ * wake→STT→tool→TTS chain into an emulator @Test is left to a follow-up
+ * PR that uses `androidx.test.rule.ServiceTestRule` (or an extracted
+ * non-service `AudioFocusController` shim). Until then, this guard at
+ * least catches DI regressions on the STT swap pattern.
+ */
+@HiltAndroidTest
+@RunWith(AndroidJUnit4::class)
+class FakeSttPipelineE2ETest {
+
+    @get:Rule val hiltRule = HiltAndroidRule(this)
+
+    @Inject lateinit var fakeStt: FakeSpeechToText
+    @Inject lateinit var realInterface: SpeechToText
+
+    @Before
+    fun setUp() {
+        hiltRule.inject()
+        fakeStt.reset()
+    }
+
+    @Test
+    fun hilt_resolves_fake_for_speech_to_text_interface() {
+        // Both injection points must resolve to the same singleton —
+        // otherwise tests that inject the interface (e.g. via a real
+        // production component) would not see queued results.
+        assertThat(realInterface).isSameInstanceAs(fakeStt)
+    }
+
+    @Test
+    fun queued_results_emit_in_order_and_terminate_on_final() = runTest {
+        fakeStt.queue(SttResult.Partial("set timer"))
+        fakeStt.queue(SttResult.Partial("set timer for 5"))
+        fakeStt.queue(SttResult.Final("set timer for 5 minutes"))
+
+        val collected = realInterface.startListening().toList()
+
+        assertThat(collected).hasSize(3)
+        assertThat(collected[0]).isEqualTo(SttResult.Partial("set timer"))
+        assertThat(collected[2]).isInstanceOf(SttResult.Final::class.java)
+        assertThat((collected[2] as SttResult.Final).text)
+            .isEqualTo("set timer for 5 minutes")
+    }
+
+    @Test
+    fun error_result_terminates_listening() = runTest {
+        fakeStt.queue(SttResult.Partial("hello"))
+        fakeStt.queue(SttResult.Error("NETWORK"))
+
+        val collected = realInterface.startListening().toList()
+
+        assertThat(collected).hasSize(2)
+        assertThat(collected[1]).isInstanceOf(SttResult.Error::class.java)
+    }
+
+    @Test
+    fun reset_allows_a_second_listening_session() = runTest {
+        fakeStt.queue(SttResult.Final("first"))
+        val first = realInterface.startListening().toList()
+        assertThat(first).hasSize(1)
+
+        // Without reset (or implicit channel renewal in startListening), a
+        // second call would see a closed channel.
+        fakeStt.queue(SttResult.Final("second"))
+        val second = realInterface.startListening().toList()
+        assertThat(second).hasSize(1)
+        assertThat((second[0] as SttResult.Final).text).isEqualTo("second")
+    }
+}
diff --git a/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSpeechToText.kt b/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSpeechToText.kt
new file mode 100644
index 00000000..454c8275
--- /dev/null
+++ b/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSpeechToText.kt
@@ -0,0 +1,70 @@
+package com.opendash.app.e2e.fakes
+
+import com.opendash.app.voice.stt.SpeechToText
+import com.opendash.app.voice.stt.SttResult
+import kotlinx.coroutines.channels.Channel
+import kotlinx.coroutines.flow.Flow
+import kotlinx.coroutines.flow.MutableStateFlow
+import kotlinx.coroutines.flow.StateFlow
+import kotlinx.coroutines.flow.asStateFlow
+import kotlinx.coroutines.flow.consumeAsFlow
+
+/**
+ * Test double for [SpeechToText] that lets a test programmatically push
+ * [SttResult]s into the pipeline without needing a real microphone or
+ * SpeechRecognizer service.
+ *
+ * Usage in a test:
+ * ```
+ * @Inject lateinit var fakeStt: FakeSpeechToText
+ * fakeStt.queue(SttResult.Final("set timer for 5 minutes"))
+ * voicePipeline.startListening()  // collects from this fake
+ * ```
+ *
+ * Each call to [startListening] returns a fresh Flow backed by a fresh
+ * channel so tests can re-enter listening mode multiple times without
+ * cross-talk between iterations.
+ */
+class FakeSpeechToText : SpeechToText {
+
+    private var channel: Channel<SttResult> = Channel(Channel.UNLIMITED)
+    private val _isListening = MutableStateFlow(false)
+    override val isListening: StateFlow<Boolean> = _isListening.asStateFlow()
+
+    /**
+     * Push an [SttResult] that the next (or current) collector will receive.
+     * Calls before [startListening] are buffered until a collector subscribes.
+     */
+    fun queue(result: SttResult) {
+        channel.trySend(result)
+        // A `Final` or `Error` result terminates the current listening session
+        // — close the channel so the collector sees the Flow complete and
+        // VoicePipeline transitions out of Listening.
+        if (result is SttResult.Final || result is SttResult.Error) {
+            channel.close()
+        }
+    }
+
+    override fun startListening(): Flow<SttResult> {
+        // Reset for a new listening session if the previous one already
+        // completed. A test may call queue → startListening → queue
+        // multiple times within one @Test method.
+        if (channel.isClosedForSend) {
+            channel = Channel(Channel.UNLIMITED)
+        }
+        _isListening.value = true
+        val current = channel
+        return current.consumeAsFlow()
+    }
+
+    override fun stopListening() {
+        _isListening.value = false
+        channel.close()
+    }
+
+    fun reset() {
+        _isListening.value = false
+        if (!channel.isClosedForSend) channel.close()
+        channel = Channel(Channel.UNLIMITED)
+    }
+}
diff --git a/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSttTestModule.kt b/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSttTestModule.kt
new file mode 100644
index 00000000..cb9affc0
--- /dev/null
+++ b/app/src/androidTest/java/com/opendash/app/e2e/fakes/FakeSttTestModule.kt
@@ -0,0 +1,36 @@
+package com.opendash.app.e2e.fakes
+
+import com.opendash.app.di.SttModule
+import com.opendash.app.voice.stt.SpeechToText
+import dagger.Module
+import dagger.Provides
+import dagger.hilt.components.SingletonComponent
+import dagger.hilt.testing.TestInstallIn
+import javax.inject.Singleton
+
+/**
+ * Replaces [SttModule] in instrumented tests so the production
+ * [com.opendash.app.voice.stt.DelegatingSttProvider] (which would talk to
+ * Android's `SpeechRecognizer`) is swapped for a programmatic
+ * [FakeSpeechToText].
+ *
+ * Tests inject `FakeSpeechToText` directly via
+ * `@Inject lateinit var fakeStt: FakeSpeechToText` — the binding below
+ * registers the same singleton against both [SpeechToText] and the
+ * concrete fake type.
+ */
+@Module
+@TestInstallIn(
+    components = [SingletonComponent::class],
+    replaces = [SttModule::class]
+)
+object FakeSttTestModule {
+
+    @Provides
+    @Singleton
+    fun provideFakeStt(): FakeSpeechToText = FakeSpeechToText()
+
+    @Provides
+    @Singleton
+    fun provideSpeechToText(fake: FakeSpeechToText): SpeechToText = fake
+}
diff --git a/app/src/main/java/com/opendash/app/di/SttModule.kt b/app/src/main/java/com/opendash/app/di/SttModule.kt
new file mode 100644
index 00000000..6a6ba3f0
--- /dev/null
+++ b/app/src/main/java/com/opendash/app/di/SttModule.kt
@@ -0,0 +1,36 @@
+package com.opendash.app.di
+
+import android.content.Context
+import com.opendash.app.data.preferences.AppPreferences
+import com.opendash.app.voice.stt.AndroidSttProvider
+import com.opendash.app.voice.stt.DelegatingSttProvider
+import com.opendash.app.voice.stt.SpeechToText
+import dagger.Module
+import dagger.Provides
+import dagger.hilt.InstallIn
+import dagger.hilt.android.qualifiers.ApplicationContext
+import dagger.hilt.components.SingletonComponent
+import javax.inject.Singleton
+
+/**
+ * SpeechToText binding lives in its own module so instrumented tests can
+ * swap the implementation via `@TestInstallIn(replaces = [SttModule::class])`
+ * without rebuilding the rest of the voice graph (TTS, VoicePipeline,
+ * LatencyRecorder).
+ *
+ * See `app/src/androidTest/.../FakeSttTestModule.kt`.
+ */
+@Module
+@InstallIn(SingletonComponent::class)
+object SttModule {
+
+    @Provides
+    @Singleton
+    fun provideSpeechToText(
+        @ApplicationContext context: Context,
+        preferences: AppPreferences
+    ): SpeechToText = DelegatingSttProvider(
+        preferences = preferences,
+        android = AndroidSttProvider(context)
+    )
+}
diff --git a/app/src/main/java/com/opendash/app/di/VoiceModule.kt b/app/src/main/java/com/opendash/app/di/VoiceModule.kt
index bb8fd5dd..a1bc7b66 100644
--- a/app/src/main/java/com/opendash/app/di/VoiceModule.kt
+++ b/app/src/main/java/com/opendash/app/di/VoiceModule.kt
@@ -10,8 +10,6 @@ import com.opendash.app.voice.fastpath.FastPathRouter
 import com.opendash.app.voice.metrics.LatencyRecorder
 import com.opendash.app.voice.pipeline.FastPathLlmPolisher
 import com.opendash.app.voice.pipeline.VoicePipeline
-import com.opendash.app.voice.stt.AndroidSttProvider
-import com.opendash.app.voice.stt.DelegatingSttProvider
 import com.opendash.app.voice.stt.SpeechToText
 import com.opendash.app.voice.tts.TextToSpeech
 import com.squareup.moshi.Moshi
@@ -26,16 +24,8 @@ import javax.inject.Singleton
 @InstallIn(SingletonComponent::class)
 object VoiceModule {
 
-    @Provides
-    @Singleton
-    fun provideSpeechToText(
-        @ApplicationContext context: Context,
-        preferences: AppPreferences
-    ): SpeechToText = DelegatingSttProvider(
-        preferences = preferences,
-        android = AndroidSttProvider(context)
-    )
-
+    // provideSpeechToText moved to SttModule so @TestInstallIn can swap STT
+    // independently of the rest of the voice graph.
     // provideTextToSpeech moved to TtsModule so @TestInstallIn can swap the
     // TTS implementation independently of the rest of the voice graph.