diff --git a/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift b/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
index bd561f02..a5d84292 100644
--- a/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
+++ b/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
@@ -28,8 +28,6 @@ final class VoiceEngineSettingsViewModel: ObservableObject {
     @Published var selectedSpeechProvider: SettingsStore.SpeechModel.Provider
     @Published var previewSpeechModel: SettingsStore.SpeechModel
     @Published var showAdvancedSpeechInfo: Bool = false
-    @Published var suppressSpeechProviderSync: Bool = false
-    @Published var skipNextSpeechModelSync: Bool = false
 
     var downloadingModel: SettingsStore.SpeechModel? {
         guard let modelID = self.asr.downloadingModelId else { return nil }
@@ -72,11 +70,6 @@ final class VoiceEngineSettingsViewModel: ObservableObject {
     }
 
     func handleSelectedSpeechModelChange(_ newValue: SettingsStore.SpeechModel) {
-        if self.skipNextSpeechModelSync {
-            self.skipNextSpeechModelSync = false
-            return
-        }
-        guard !self.suppressSpeechProviderSync else { return }
         self.previewSpeechModel = newValue
         self.setSelectedSpeechProvider(newValue.provider)
     }
@@ -172,32 +165,19 @@ final class VoiceEngineSettingsViewModel: ObservableObject {
 
     func deleteSpeechModel(_ model: SettingsStore.SpeechModel) {
         guard !self.areSpeechModelActionsBlocked else { return }
-        let previousActive = self.settings.selectedSpeechModel
-
         Task {
-            let shouldRestore = previousActive != model
-            await MainActor.run {
-                if shouldRestore {
-                    self.suppressSpeechProviderSync = true
-                }
-                self.settings.selectedSpeechModel = model
-                self.asr.resetTranscriptionProvider()
-            }
-
-            defer {
-                Task { @MainActor in
-                    guard shouldRestore else { return }
-                    self.skipNextSpeechModelSync = true
-                    self.settings.selectedSpeechModel = previousActive
-                    self.asr.resetTranscriptionProvider()
-                    if self.previewSpeechModel == model {
-                        self.previewSpeechModel = model
-                    }
-                    self.suppressSpeechProviderSync = false
-                }
+            do {
+                try await self.asr.clearModelCache(for: model)
+                // Refresh installed-model state so the model cards reflect the deletion for
+                // every model, not only the active one (clearModelCache(for:) re-checks only
+                // when deleting the selected model).
+                await self.asr.checkIfModelsExistAsync()
+            } catch {
+                DebugLogger.shared.error(
+                    "Failed to delete model \(model.displayName): \(error)",
+                    source: "VoiceEngineVM"
+                )
             }
-
-            await self.deleteModels()
         }
     }
 
diff --git a/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift b/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
index d659e03b..e187bb2b 100644
--- a/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
+++ b/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
@@ -359,6 +359,56 @@ final class DictationE2ETests: XCTestCase {
         )
     }
 
+    func testDeleteNonActiveSpeechModel_neverPersistsDeletedModelAsSelection() async throws {
+        // Regression: deleting a NON-ACTIVE speech model used to temporarily persist the
+        // deleted model as `selectedSpeechModel`, then restore the real selection through an
+        // un-awaited `defer { Task { ... } }`. If the process exited during that window (app
+        // quit, crash, or update relaunch) the user was left with a model they never chose.
+        // The fix deletes the target model's cache without ever writing the selection.
+        let defaults = UserDefaults.standard
+        let key = "SelectedSpeechModel"
+        let snapshot = defaults.object(forKey: key)
+        defer {
+            if let snapshot {
+                defaults.set(snapshot, forKey: key)
+            } else {
+                defaults.removeObject(forKey: key)
+            }
+        }
+
+        let settings = SettingsStore.shared
+        let activeModel: SettingsStore.SpeechModel = .whisperTiny
+        let modelToDelete: SettingsStore.SpeechModel = .appleSpeech
+        XCTAssertNotEqual(activeModel, modelToDelete)
+
+        settings.selectedSpeechModel = activeModel
+        XCTAssertEqual(settings.selectedSpeechModel, activeModel)
+
+        // Record every value the persisted selection takes during the delete so a transient
+        // write of the deleted model is caught directly, not masked by the async restore.
+        let recorder = PersistedSpeechModelRecorder(defaults: defaults, key: key)
+        defer { recorder.stop() }
+
+        let viewModel = VoiceEngineSettingsViewModel(settings: settings, appServices: AppServices.shared)
+        viewModel.deleteSpeechModel(modelToDelete)
+
+        // `deleteSpeechModel` is fire-and-forget; let its async work and any deferred restore settle.
+        try await Task.sleep(nanoseconds: 1_000_000_000)
+
+        XCTAssertEqual(
+            settings.selectedSpeechModel,
+            activeModel,
+            "Deleting a non-active model must leave the user's selection unchanged."
+        )
+        XCTAssertFalse(
+            recorder.recordedRawValues.contains(modelToDelete.rawValue),
+            """
+            Deleting a non-active model must never persist it as the selection, even transiently. \
+            Recorded writes: \(recorder.recordedRawValues)
+            """
+        )
+    }
+
     func testAppPromptBinding_profileOverridesModeSelection() {
         self.withPromptSettingsRestored {
             let settings = SettingsStore.shared
@@ -1055,3 +1105,40 @@ final class DictationE2ETests: XCTestCase {
         )
     }
 }
+
+// Classic key-path KVO is required here: an arbitrary UserDefaults key has no Swift `KeyPath`
+// for the block-based API, and the change dictionary is delivered synchronously with the exact
+// new value, which a coalescing `UserDefaults.didChangeNotification` cannot guarantee.
+// swiftlint:disable block_based_kvo discouraged_optional_collection
+
+/// Records every value written to a UserDefaults key via KVO, so a transient write that is
+/// later reverted can still be observed. Used to catch the non-active speech-model delete race,
+/// where the deleted model was momentarily persisted as the selection before being restored.
+private final class PersistedSpeechModelRecorder: NSObject {
+    private let defaults: UserDefaults
+    private let key: String
+    private(set) var recordedRawValues: [String] = []
+
+    init(defaults: UserDefaults, key: String) {
+        self.defaults = defaults
+        self.key = key
+        super.init()
+        defaults.addObserver(self, forKeyPath: key, options: [.new], context: nil)
+    }
+
+    func stop() {
+        self.defaults.removeObserver(self, forKeyPath: self.key)
+    }
+
+    override func observeValue(
+        forKeyPath keyPath: String?,
+        of object: Any?,
+        change: [NSKeyValueChangeKey: Any]?,
+        context: UnsafeMutableRawPointer?
+    ) {
+        guard keyPath == self.key, let rawValue = change?[.newKey] as? String else { return }
+        self.recordedRawValues.append(rawValue)
+    }
+}
+
+// swiftlint:enable block_based_kvo discouraged_optional_collection