Skip to content

Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more

feat(asr): migrate Whisper to TranscribeCpp

altic-deva7876b16cc3f
diff --git a/Fluid.xcodeproj/project.pbxproj b/Fluid.xcodeproj/project.pbxproj
index 3c98aacd..a2cb0a9f 100644
--- a/Fluid.xcodeproj/project.pbxproj
+++ b/Fluid.xcodeproj/project.pbxproj
@@ -8,9 +8,9 @@
/* Begin PBXBuildFile section */
7C078DA52E3B386A00FB7CAC /* FluidAudio in Frameworks */ = {isa = PBXBuildFile; productRef = 7C078DA42E3B386A00FB7CAC /* FluidAudio */; };
- 7C1C72F22EECBD1300E3BF4D /* SwiftWhisper in Frameworks */ = {isa = PBXBuildFile; productRef = 7C1C72F12EECBD1300E3BF4D /* SwiftWhisper */; };
7C3697892ED70F9C005874CE /* DynamicNotchKit in Frameworks */ = {isa = PBXBuildFile; productRef = 7C3697882ED70F9C005874CE /* DynamicNotchKit */; };
7C5AF14B2F15041600DE21B0 /* MediaRemoteAdapter in Frameworks */ = {isa = PBXBuildFile; productRef = 7C5AF14A2F15041600DE21B0 /* MediaRemoteAdapter */; };
+ 7C9A71022F58B00000FB7CAF /* TranscribeCpp in Frameworks */ = {isa = PBXBuildFile; productRef = 7C9A71012F58B00000FB7CAF /* TranscribeCpp */; };
7C91B0012F42AA0100C0DEF0 /* HotkeyShortcutTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7C91B0022F42AA0100C0DEF0 /* HotkeyShortcutTests.swift */; };
7CDB0A2D2F3C4D5600FB7CAD /* DictationE2ETests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7CDB0A292F3C4D5600FB7CAD /* DictationE2ETests.swift */; };
7CDB0A2E2F3C4D5600FB7CAD /* AudioFixtureLoader.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7CDB0A2A2F3C4D5600FB7CAD /* AudioFixtureLoader.swift */; };
@@ -58,7 +58,7 @@
files = (
7C3697892ED70F9C005874CE /* DynamicNotchKit in Frameworks */,
7C5AF14B2F15041600DE21B0 /* MediaRemoteAdapter in Frameworks */,
- 7C1C72F22EECBD1300E3BF4D /* SwiftWhisper in Frameworks */,
+ 7C9A71022F58B00000FB7CAF /* TranscribeCpp in Frameworks */,
7CE006BD2E80EBE600DDCCD6 /* AppUpdater in Frameworks */,
7C078DA52E3B386A00FB7CAC /* FluidAudio in Frameworks */,
);
@@ -161,7 +161,7 @@
7C078DA42E3B386A00FB7CAC /* FluidAudio */,
7C3697882ED70F9C005874CE /* DynamicNotchKit */,
7CE006BC2E80EBE600DDCCD6 /* AppUpdater */,
- 7C1C72F12EECBD1300E3BF4D /* SwiftWhisper */,
+ 7C9A71012F58B00000FB7CAF /* TranscribeCpp */,
7C5AF14A2F15041600DE21B0 /* MediaRemoteAdapter */,
);
productName = FluidVoice;
@@ -217,7 +217,7 @@
7C078DA32E3B386A00FB7CAC /* XCRemoteSwiftPackageReference "FluidAudio" */,
7CE006BB2E80EBE600DDCCD6 /* XCRemoteSwiftPackageReference "AppUpdater" */,
7C3697872ED70F9C005874CE /* XCRemoteSwiftPackageReference "DynamicNotchKit" */,
- 7C1C72F02EECBD1300E3BF4D /* XCRemoteSwiftPackageReference "SwiftWhisper" */,
+ 7C9A71002F58B00000FB7CAF /* XCRemoteSwiftPackageReference "transcribe-cpp-swift" */,
7C5AF1492F15041600DE21B0 /* XCRemoteSwiftPackageReference "mediaremote-adapter" */,
);
preferredProjectObjectVersion = 77;
@@ -605,14 +605,6 @@
kind = branch;
};
};
- 7C1C72F02EECBD1300E3BF4D /* XCRemoteSwiftPackageReference "SwiftWhisper" */ = {
- isa = XCRemoteSwiftPackageReference;
- repositoryURL = "https://github.com/exPHAT/SwiftWhisper.git";
- requirement = {
- kind = upToNextMajorVersion;
- minimumVersion = 1.2.0;
- };
- };
7C3697872ED70F9C005874CE /* XCRemoteSwiftPackageReference "DynamicNotchKit" */ = {
isa = XCRemoteSwiftPackageReference;
repositoryURL = "https://github.com/altic-dev/DynamicNotchKit.git";
@@ -629,6 +621,14 @@
kind = branch;
};
};
+ 7C9A71002F58B00000FB7CAF /* XCRemoteSwiftPackageReference "transcribe-cpp-swift" */ = {
+ isa = XCRemoteSwiftPackageReference;
+ repositoryURL = "https://github.com/altic-dev/transcribe-cpp-swift.git";
+ requirement = {
+ kind = exactVersion;
+ version = 0.1.2;
+ };
+ };
7CE006BB2E80EBE600DDCCD6 /* XCRemoteSwiftPackageReference "AppUpdater" */ = {
isa = XCRemoteSwiftPackageReference;
repositoryURL = "https://github.com/mxcl/AppUpdater.git";
@@ -645,11 +645,6 @@
package = 7C078DA32E3B386A00FB7CAC /* XCRemoteSwiftPackageReference "FluidAudio" */;
productName = FluidAudio;
};
- 7C1C72F12EECBD1300E3BF4D /* SwiftWhisper */ = {
- isa = XCSwiftPackageProductDependency;
- package = 7C1C72F02EECBD1300E3BF4D /* XCRemoteSwiftPackageReference "SwiftWhisper" */;
- productName = SwiftWhisper;
- };
7C3697882ED70F9C005874CE /* DynamicNotchKit */ = {
isa = XCSwiftPackageProductDependency;
package = 7C3697872ED70F9C005874CE /* XCRemoteSwiftPackageReference "DynamicNotchKit" */;
@@ -665,6 +660,11 @@
package = 7CE006BB2E80EBE600DDCCD6 /* XCRemoteSwiftPackageReference "AppUpdater" */;
productName = AppUpdater;
};
+ 7C9A71012F58B00000FB7CAF /* TranscribeCpp */ = {
+ isa = XCSwiftPackageProductDependency;
+ package = 7C9A71002F58B00000FB7CAF /* XCRemoteSwiftPackageReference "transcribe-cpp-swift" */;
+ productName = TranscribeCpp;
+ };
/* End XCSwiftPackageProductDependency section */
};
rootObject = 7C078D872E3B339200FB7CAC /* Project object */;
diff --git a/Fluid.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved b/Fluid.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved
index dbfbbd53..85916ad3 100644
--- a/Fluid.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved
+++ b/Fluid.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved
@@ -1,5 +1,5 @@
{
- "originHash" : "064855d16f7e949cfc68503de53276b1c6192736444e73922fc77d0a6b315471",
+ "originHash" : "c0b937cf1a9f455cc47d65c349124c6a9e9f173f41a1aa7809c1c0d65cbf4229",
"pins" : [
{
"identity" : "appupdater",
@@ -128,12 +128,12 @@
}
},
{
- "identity" : "swiftwhisper",
+ "identity" : "transcribe-cpp-swift",
"kind" : "remoteSourceControl",
- "location" : "https://github.com/exPHAT/SwiftWhisper.git",
+ "location" : "https://github.com/altic-dev/transcribe-cpp-swift.git",
"state" : {
- "revision" : "a192004db08de7c6eaa169eede77f1625e7d23fb",
- "version" : "1.2.0"
+ "revision" : "a16df2905f9c715344374fd2d8dd97c2712a6d96",
+ "version" : "0.1.2"
}
},
{
diff --git a/Package.resolved b/Package.resolved
index 33ebe338..f5f21e17 100644
--- a/Package.resolved
+++ b/Package.resolved
@@ -91,12 +91,12 @@
}
},
{
- "identity" : "swiftwhisper",
+ "identity" : "transcribe-cpp-swift",
"kind" : "remoteSourceControl",
- "location" : "https://github.com/exPHAT/SwiftWhisper.git",
+ "location" : "https://github.com/altic-dev/transcribe-cpp-swift.git",
"state" : {
- "branch" : "master",
- "revision" : "c340197966ebd264f3135d3955874b40f8ed58bc"
+ "revision" : "a16df2905f9c715344374fd2d8dd97c2712a6d96",
+ "version" : "0.1.2"
}
},
{
diff --git a/Package.swift b/Package.swift
index 740ac232..a680e275 100644
--- a/Package.swift
+++ b/Package.swift
@@ -6,14 +6,14 @@ import PackageDescription
let package = Package(
name: "FluidVoice",
platforms: [
- .macOS(.v15),
+ .macOS("15.0"),
],
dependencies: [
.package(url: "https://github.com/mxcl/AppUpdater.git", from: "1.0.0"),
.package(url: "https://github.com/altic-dev/FluidAudio.git", branch: "B/cohere-coreml-asr"),
.package(url: "https://github.com/mxcl/PromiseKit", from: "6.0.0"),
.package(url: "https://github.com/altic-dev/DynamicNotchKit.git", branch: "main"),
- .package(url: "https://github.com/exPHAT/SwiftWhisper.git", branch: "master"),
+ .package(url: "https://github.com/altic-dev/transcribe-cpp-swift.git", exact: "0.1.2"),
.package(url: "https://github.com/PostHog/posthog-ios.git", from: "3.0.0"),
],
targets: [
@@ -32,7 +32,7 @@ let package = Package(
"FluidAudio",
"PromiseKit",
"DynamicNotchKit",
- "SwiftWhisper",
+ .product(name: "TranscribeCpp", package: "transcribe-cpp-swift"),
.product(name: "PostHog", package: "posthog-ios"),
]
),
diff --git a/Sources/Fluid/AppDelegate.swift b/Sources/Fluid/AppDelegate.swift
index b70fa9f2..d1b89205 100644
--- a/Sources/Fluid/AppDelegate.swift
+++ b/Sources/Fluid/AppDelegate.swift
@@ -65,12 +65,33 @@ class AppDelegate: NSObject, NSApplicationDelegate, UNUserNotificationCenterDele
func applicationWillTerminate(_ notification: Notification) {
DebugLogger.shared.info("Application will terminate", source: "AppDelegate")
self.shutdownPrivateAIRuntimeForTermination()
+ self.shutdownASRRuntimeForTermination()
LocalAPIServer.shared.stop()
// Clean up the update check timer
self.updateCheckTimer?.invalidate()
self.updateCheckTimer = nil
}
+ private func shutdownASRRuntimeForTermination() {
+ var didFinishShutdown = false
+ Task { @MainActor in
+ await AppServices.shared.shutdownForTermination()
+ didFinishShutdown = true
+ }
+
+ let deadline = Date().addingTimeInterval(8)
+ while !didFinishShutdown, Date() < deadline {
+ RunLoop.current.run(mode: .default, before: Date().addingTimeInterval(0.01))
+ }
+
+ if !didFinishShutdown {
+ DebugLogger.shared.warning(
+ "Timed out waiting for ASR runtime shutdown during termination",
+ source: "AppDelegate"
+ )
+ }
+ }
+
private func shutdownPrivateAIRuntimeForTermination() {
var didFinishShutdown = false
Task { @MainActor in
diff --git a/Sources/Fluid/Persistence/SettingsStore.swift b/Sources/Fluid/Persistence/SettingsStore.swift
index 46313e66..35340cd7 100644
--- a/Sources/Fluid/Persistence/SettingsStore.swift
+++ b/Sources/Fluid/Persistence/SettingsStore.swift
@@ -3939,7 +3939,7 @@ final class SettingsStore: ObservableObject {
case whisperBase = "whisper-base"
case whisperSmall = "whisper-small"
case whisperMedium = "whisper-medium"
- case whisperLargeTurbo = "whisper-large-turbo" // temporarily disabled in UI
+ case whisperLargeTurbo = "whisper-large-turbo"
case whisperLarge = "whisper-large"
var id: String {
@@ -3964,7 +3964,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return "Whisper Base"
case .whisperSmall: return "Whisper Small"
case .whisperMedium: return "Whisper Medium"
- case .whisperLargeTurbo: return "Whisper Large Turbo (Disabled)"
+ case .whisperLargeTurbo: return "Whisper Large Turbo"
case .whisperLarge: return "Whisper Large"
}
}
@@ -3997,12 +3997,12 @@ final class SettingsStore: ObservableObject {
case .nemotronStreaming320: return "~668.2 MiB"
case .appleSpeech: return "Built-in"
case .appleSpeechAnalyzer: return "Built-in"
- case .whisperTiny: return "~74.1 MiB"
- case .whisperBase: return "~141.1 MiB"
- case .whisperSmall: return "~465.0 MiB"
- case .whisperMedium: return "~1.43 GiB"
- case .whisperLargeTurbo: return "~1.51 GiB"
- case .whisperLarge: return "~2.88 GiB"
+ case .whisperTiny: return "~43.9 MiB"
+ case .whisperBase: return "~81.0 MiB"
+ case .whisperSmall: return "~257.3 MiB"
+ case .whisperMedium: return "~793.0 MiB"
+ case .whisperLargeTurbo: return "~845.3 MiB"
+ case .whisperLarge: return "~1.55 GiB"
}
}
@@ -4015,12 +4015,12 @@ final class SettingsStore: ObservableObject {
case .cohereTranscribeSixBit: return 1_650_748_785
case .nemotronOffline: return 556_552_620
case .nemotronStreaming, .nemotronStreaming320: return 700_685_415
- case .whisperTiny: return 77_691_713
- case .whisperBase: return 147_951_465
- case .whisperSmall: return 487_601_967
- case .whisperMedium: return 1_533_763_059
- case .whisperLargeTurbo: return 1_624_555_275
- case .whisperLarge: return 3_095_033_483
+ case .whisperTiny: return 45_981_088
+ case .whisperBase: return 84_962_880
+ case .whisperSmall: return 269_751_136
+ case .whisperMedium: return 831_538_144
+ case .whisperLargeTurbo: return 886_381_824
+ case .whisperLarge: return 1_668_741_440
case .appleSpeech, .appleSpeechAnalyzer: return 0
}
}
@@ -4039,8 +4039,20 @@ final class SettingsStore: ObservableObject {
}
}
- /// The ggml filename for Whisper models
+ /// The GGUF filename for transcribe.cpp Whisper models.
var whisperModelFile: String? {
+ switch self {
+ case .whisperTiny: return "whisper-tiny-Q8_0.gguf"
+ case .whisperBase: return "whisper-base-Q8_0.gguf"
+ case .whisperSmall: return "whisper-small-Q8_0.gguf"
+ case .whisperMedium: return "whisper-medium-Q8_0.gguf"
+ case .whisperLargeTurbo: return "whisper-large-v3-turbo-Q8_0.gguf"
+ case .whisperLarge: return "whisper-large-v3-Q8_0.gguf"
+ default: return nil
+ }
+ }
+
+ var legacyWhisperModelFile: String? {
switch self {
case .whisperTiny: return "ggml-tiny.bin"
case .whisperBase: return "ggml-base.bin"
@@ -4052,6 +4064,15 @@ final class SettingsStore: ObservableObject {
}
}
+ static let legacyWhisperModelFiles: Set<String> = [
+ "ggml-tiny.bin",
+ "ggml-base.bin",
+ "ggml-small.bin",
+ "ggml-medium.bin",
+ "ggml-large-v3-turbo.bin",
+ "ggml-large-v3.bin",
+ ]
+
/// The short model name for whisper.cpp internal usage
var whisperModelName: String? {
switch self {
@@ -4086,7 +4107,10 @@ final class SettingsStore: ObservableObject {
/// Returns models available for the current Mac's architecture and OS
static var availableModels: [SpeechModel] {
allCases.filter { model in
- if model == .whisperLargeTurbo {
+ if model == .whisperLargeTurbo, !CPUArchitecture.isAppleSilicon {
+ return false
+ }
+ if model == .whisperLarge, !CPUArchitecture.isAppleSilicon {
return false
}
if model == .qwen3Asr, !Self.qwenPreviewEnabled {
@@ -4205,11 +4229,11 @@ final class SettingsStore: ObservableObject {
case .whisperSmall:
return 4.0
case .whisperMedium:
- return 6.0
+ return 5.0
case .whisperLargeTurbo:
- return 8.0
+ return 6.0
case .whisperLarge:
- return 10.0 // Large model needs ~6-8GB working memory + model size
+ return 8.0
}
}
@@ -4976,6 +5000,10 @@ extension SettingsStore {
if model == .nemotronStreaming320 {
return .nemotronStreaming
}
+ let requiresAppleSiliconWhisper = model == .whisperLargeTurbo || model == .whisperLarge
+ if requiresAppleSiliconWhisper, !CPUArchitecture.isAppleSilicon {
+ return .whisperBase
+ }
// Validate model is available on this architecture
if model.requiresAppleSilicon && !CPUArchitecture.isAppleSilicon {
return .whisperBase
@@ -5072,7 +5100,7 @@ extension SettingsStore {
case "ggml-base.bin": newModel = .whisperBase
case "ggml-small.bin": newModel = .whisperSmall
case "ggml-medium.bin": newModel = .whisperMedium
- case "ggml-large-v3.bin": newModel = .whisperLarge
+ case "ggml-large-v3.bin": newModel = CPUArchitecture.isAppleSilicon ? .whisperLarge : .whisperBase
default: newModel = .whisperBase
}
case "fluidAudio":
diff --git a/Sources/Fluid/Persistence/VoiceEngineLanguageCatalog.swift b/Sources/Fluid/Persistence/VoiceEngineLanguageCatalog.swift
index 5608d10c..7d380539 100644
--- a/Sources/Fluid/Persistence/VoiceEngineLanguageCatalog.swift
+++ b/Sources/Fluid/Persistence/VoiceEngineLanguageCatalog.swift
@@ -2,7 +2,6 @@ import Foundation
#if canImport(Speech)
import Speech
#endif
-import SwiftWhisper
struct VoiceEngineLanguage: Identifiable, Equatable {
let id: String
@@ -350,21 +349,28 @@ enum VoiceEngineLanguageCatalog {
}
private static let whisperLanguageCodeMap: [String: String] = {
- let supportedCodes = Set(
- WhisperLanguage.allCases
- .map(\.rawValue)
- .filter { $0 != WhisperLanguage.auto.rawValue }
- )
-
var languageMap: [String: String] = [:]
for language in Self.languageDefinitions {
let whisperCode = language.id == "he" ? "iw" : language.id
- guard supportedCodes.contains(whisperCode) else { continue }
+ guard Self.whisperSupportedLanguageCodes.contains(whisperCode) else { continue }
languageMap[language.id] = whisperCode
}
return languageMap
}()
+ private static let whisperSupportedLanguageCodes: Set<String> = [
+ "af", "am", "ar", "as", "az", "ba", "be", "bg", "bn", "bo",
+ "br", "bs", "ca", "cs", "cy", "da", "de", "el", "en", "es",
+ "et", "eu", "fa", "fi", "fo", "fr", "gl", "gu", "ha", "haw",
+ "hi", "hr", "ht", "hu", "hy", "id", "is", "it", "iw", "ja",
+ "jw", "ka", "kk", "km", "kn", "ko", "la", "lb", "ln", "lo",
+ "lt", "lv", "mg", "mi", "mk", "ml", "mn", "mr", "ms", "mt",
+ "my", "ne", "nl", "nn", "no", "oc", "pa", "pl", "ps", "pt",
+ "ro", "ru", "sa", "sd", "si", "sk", "sl", "sn", "so", "sq",
+ "sr", "su", "sv", "sw", "ta", "te", "tg", "th", "tk", "tl",
+ "tr", "tt", "uk", "ur", "uz", "vi", "yi", "yo", "zh",
+ ]
+
private static let languageDefinitions: [VoiceEngineLanguage] = [
Self.language("af", "Afrikaans"),
Self.language("am", "Amharic"),
diff --git a/Sources/Fluid/Services/ASRService.swift b/Sources/Fluid/Services/ASRService.swift
index 04873922..6735e281 100644
--- a/Sources/Fluid/Services/ASRService.swift
+++ b/Sources/Fluid/Services/ASRService.swift
@@ -10,34 +10,32 @@ import AppKit
import AudioToolbox
import CoreAudio
-/// Serializes all CoreML transcription operations to prevent concurrent access issues.
-/// The actor ensures only one transcription runs at a time, preventing CoreML race conditions.
-/// Serializes all CoreML transcription operations to prevent concurrent access issues.
-/// This implementation enforces strict serialization (non-reentrant) using a task chain.
+/// Serializes transcription operations and lets teardown cancel the real queued work.
private actor TranscriptionExecutor {
private var lastTask: Task<Void, Never>?
- private var currentOperationTask: Task<Any, Error>?
+ private var operationCancellations: [UUID: () -> Void] = [:]
func run<T>(_ operation: @escaping () async throws -> T) async throws -> T {
let previous = self.lastTask
+ let operationID = UUID()
let task = Task<T, Error> {
_ = await previous?.result
+ try Task.checkCancellation()
return try await operation()
}
- self.currentOperationTask = Task<Any, Error> { try await task.value }
+ self.operationCancellations[operationID] = { task.cancel() }
self.lastTask = Task { _ = try? await task.value }
+ defer { self.operationCancellations.removeValue(forKey: operationID) }
return try await task.value
}
- /// Cancels any pending operations and waits for the current chain to complete.
- /// This ensures no in-flight transcription tasks can access deallocated memory.
func cancelAndAwaitPending() async {
- // Cancel the current operation if running
- self.currentOperationTask?.cancel()
- // Wait for the last task in the chain to complete (or be cancelled)
+ for cancel in self.operationCancellations.values {
+ cancel()
+ }
_ = await self.lastTask?.result
self.lastTask = nil
- self.currentOperationTask = nil
+ self.operationCancellations.removeAll()
}
}
@@ -186,6 +184,32 @@ final class ASRService: ObservableObject {
task.cancel()
}
+ func shutdownForTermination() async {
+ if self.isRunning {
+ await self.stopWithoutTranscription()
+ }
+
+ let preparationTask = self.ensureReadyTask
+ let downloadTask = self.modelDownloadTask
+ preparationTask?.cancel()
+ downloadTask?.cancel()
+ _ = await preparationTask?.result
+ _ = await downloadTask?.result
+ await self.providerResetDrain?.task.value
+ await self.transcriptionExecutor.cancelAndAwaitPending()
+
+ self.fluidAudioProvider = nil
+ self.parakeetRealtimeProvider = nil
+ self.externalCoreMLProvider = nil
+ self.nemotronProviders.removeAll()
+ self.whisperProvider = nil
+ self.appleSpeechProvider = nil
+ self._appleSpeechAnalyzerProvider = nil
+ self.isAsrReady = false
+ self.isLoadingModel = false
+ self.isDownloadingModel = false
+ }
+
/// The transcription provider, selected based on the unified SpeechModel setting.
/// Uses the new SettingsStore.selectedSpeechModel instead of old TranscriptionProviderOption.
private var transcriptionProvider: TranscriptionProvider {
@@ -514,6 +538,10 @@ final class ASRService: ObservableObject {
self.isCancellingModelPreparation = true
task.cancel()
}
+ let resetDrainID = UUID()
+ let executor = self.transcriptionExecutor
+ let resetDrainTask = Task { await executor.cancelAndAwaitPending() }
+ self.providerResetDrain = (resetDrainID, resetDrainTask)
// Keep the task handle until its provider has stopped and cancellation cleanup has
// completed. The next ensureAsrReady call waits for it before touching the same cache.
self.ensureReadyProviderKey = nil
@@ -867,6 +895,7 @@ final class ASRService: ObservableObject {
private var streamingChunkAnalyticsSuccessCount: Int = 0
private var lastStreamingChunkFailureAnalyticsAt: Date?
private let transcriptionExecutor = TranscriptionExecutor() // Serializes all CoreML access
+ private var providerResetDrain: (id: UUID, task: Task<Void, Never>)?
private var engineConfigurationChangeObserver: NSObjectProtocol?
private var audioRouteRecoveryTask: Task<Void, Never>?
private let audioRouteRecoveryDelayNanoseconds: UInt64 = 1_000_000_000
@@ -2630,28 +2659,10 @@ final class ASRService: ObservableObject {
// Audio tap processing is handled by AudioCapturePipeline (thread-safe).
- /// Ensures that ASR models are downloaded and ready for transcription.
- ///
- /// This method handles the complete model lifecycle using the appropriate
- /// TranscriptionProvider based on CPU architecture:
- /// - Apple Silicon: FluidAudio (CoreML optimized)
- /// - Intel: SwiftWhisper (whisper.cpp)
- ///
- /// ## Performance
- /// - First run will download models (~100-500MB depending on provider)
- /// - Subsequent runs use cached models (much faster)
- /// - Model loading happens asynchronously to avoid blocking UI
- ///
- /// ## Errors
- /// Throws if model download or loading fails. Common causes:
- /// - Network connectivity issues
- /// - Insufficient disk space
func ensureAsrReady() async throws {
try await self.ensureAsrReady(progressHandler: nil)
}
- /// Ensures ASR models are ready, with an optional external progress handler.
- /// - Parameter progressHandler: Optional callback for download progress (0.0 to 1.0)
func ensureAsrReady(progressHandler: ((Double) -> Void)?) async throws {
guard self.modelDownloadTask == nil else {
throw NSError(
@@ -2660,6 +2671,12 @@ final class ASRService: ObservableObject {
userInfo: [NSLocalizedDescriptionKey: "Another model download is already in progress."]
)
}
+ if let drain = self.providerResetDrain {
+ await drain.task.value
+ if self.providerResetDrain?.id == drain.id {
+ self.providerResetDrain = nil
+ }
+ }
let provider = self.transcriptionProvider
let model = SettingsStore.shared.selectedSpeechModel
let providerKey = "\(model.id):\(type(of: provider)):\(provider.name)"
@@ -3052,6 +3069,7 @@ final class ASRService: ObservableObject {
func clearModelCache() async throws {
DebugLogger.shared.debug("Clearing model cache via transcription provider", source: "ASRService")
+ await self.transcriptionExecutor.cancelAndAwaitPending()
try await self.transcriptionProvider.clearCache()
self.isAsrReady = false
self.modelsExistOnDisk = false
@@ -3059,6 +3077,9 @@ final class ASRService: ObservableObject {
func clearModelCache(for model: SettingsStore.SpeechModel) async throws {
DebugLogger.shared.debug("Clearing model cache for \(model.displayName)", source: "ASRService")
+ if SettingsStore.shared.selectedSpeechModel == model {
+ await self.transcriptionExecutor.cancelAndAwaitPending()
+ }
let provider = self.getProvider(for: model)
try await provider.clearCache()
diff --git a/Sources/Fluid/Services/AppServices.swift b/Sources/Fluid/Services/AppServices.swift
index 0c73e026..5ceea8da 100644
--- a/Sources/Fluid/Services/AppServices.swift
+++ b/Sources/Fluid/Services/AppServices.swift
@@ -113,4 +113,11 @@ final class AppServices: ObservableObject {
DebugLogger.shared.info("✅ All services initialized", source: "AppServices")
}
+
+ func shutdownForTermination() async {
+ if let asr = self._asr {
+ await asr.shutdownForTermination()
+ self._asr = nil
+ }
+ }
}
diff --git a/Sources/Fluid/Services/TranscriptionProvider.swift b/Sources/Fluid/Services/TranscriptionProvider.swift
index a6bf7c9e..d098f2dd 100644
--- a/Sources/Fluid/Services/TranscriptionProvider.swift
+++ b/Sources/Fluid/Services/TranscriptionProvider.swift
@@ -56,7 +56,7 @@ struct ASRTranscriptionResult {
// MARK: - Transcription Provider Protocol
/// Protocol that abstracts speech-to-text transcription.
-/// Implementations can use different backends (FluidAudio, SwiftWhisper, etc.)
+/// Implementations can use different backends (FluidAudio, transcribe.cpp, etc.)
protocol TranscriptionProvider {
/// Display name of the provider
var name: String { get }
diff --git a/Sources/Fluid/Services/WhisperProvider.swift b/Sources/Fluid/Services/WhisperProvider.swift
index 948b003c..586043ea 100644
--- a/Sources/Fluid/Services/WhisperProvider.swift
+++ b/Sources/Fluid/Services/WhisperProvider.swift
@@ -1,26 +1,31 @@
import Foundation
-import SwiftWhisper
+import TranscribeCpp
-/// TranscriptionProvider implementation using SwiftWhisper (whisper.cpp) for Intel Macs.
-/// This provides on-device speech recognition that works on Intel x86_64 architecture.
+/// TranscriptionProvider implementation using transcribe.cpp for Whisper GGUF models.
final class WhisperProvider: TranscriptionProvider {
- let name = "Whisper (Intel/Universal)"
+ let name = "Whisper (Universal)"
- /// Whether this provider is supported on the current system.
- /// SwiftWhisper (whisper.cpp) works on both Intel and Apple Silicon.
var isAvailable: Bool {
- true
+ guard case .success = Self.backendInitialization else { return false }
+ if CPUArchitecture.isAppleSilicon {
+ return Transcribe.backendAvailable(.metal)
+ }
+ return Transcribe.backendAvailable(.cpu)
+ }
+
+ private static let backendInitialization: Result<Void, Error> = Result {
+ try Transcribe.initBackends()
}
- private var whisper: Whisper?
- private(set) var isReady: Bool = false
+ private let stateLock = NSLock()
+ private var model: Model?
+ private var session: Session?
+ private var ready = false
private var loadedModelName: String?
private let overriddenModelDirectory: URL?
private let urlSession: URLSession
- /// Optional model override - if set, uses this model instead of the global setting.
- /// Used for downloading specific models without changing the active selection.
var modelOverride: SettingsStore.SpeechModel?
init(modelDirectory: URL? = nil, urlSession: URLSession = .shared, modelOverride: SettingsStore.SpeechModel? = nil) {
@@ -29,75 +34,143 @@ final class WhisperProvider: TranscriptionProvider {
self.modelOverride = modelOverride
}
- /// Model filename to use - reads from override first, then unified SpeechModel setting
- /// Models: tiny (~75MB), base (~142MB), small (~466MB), medium (~1.5GB), large (~2.9GB)
+ deinit {
+ self.unloadModel()
+ }
+
+ var isReady: Bool {
+ self.stateLock.lock()
+ defer { self.stateLock.unlock() }
+ return self.ready
+ }
+
+ private var selectedModel: SettingsStore.SpeechModel {
+ self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
+ }
+
private var modelName: String {
- let model = self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
- let configured = model.whisperModelFile?
- .trimmingCharacters(in: .whitespacesAndNewlines)
- if let configured, !configured.isEmpty {
- return configured
- }
- return "ggml-base.bin"
+ self.selectedModel.whisperModelFile ?? "whisper-base-Q8_0.gguf"
+ }
+
+ private var legacyModelName: String? {
+ self.selectedModel.legacyWhisperModelFile
}
private var modelURL: URL {
- let directory: URL
+ self.modelDirectory.appendingPathComponent(self.modelName)
+ }
+
+ private var legacyModelURL: URL? {
+ self.legacyModelName.map { self.modelDirectory.appendingPathComponent($0) }
+ }
+
+ private var modelDirectory: URL {
if let overriddenModelDirectory {
- directory = overriddenModelDirectory
- } else {
- guard let cacheDir = FileManager.default.urls(for: .cachesDirectory, in: .userDomainMask).first else {
- preconditionFailure("Could not find caches directory")
- }
- directory = cacheDir.appendingPathComponent("WhisperModels")
+ return overriddenModelDirectory
}
+ guard let cacheDir = FileManager.default.urls(for: .cachesDirectory, in: .userDomainMask).first else {
+ preconditionFailure("Could not find caches directory")
+ }
+ return cacheDir.appendingPathComponent("WhisperModels")
+ }
- return directory.appendingPathComponent(self.modelName)
+ private var modelDownloadURL: URL? {
+ let modelName = self.modelName
+ let suffix = "-Q8_0.gguf"
+ guard modelName.hasSuffix(suffix) else { return nil }
+ let repoName = String(modelName.dropLast(suffix.count))
+ return URL(string: "https://huggingface.co/handy-computer/\(repoName)-gguf/resolve/main/\(modelName)")
}
- private var modelDirectory: URL {
- self.modelURL.deletingLastPathComponent()
+ private var backend: Backend {
+ CPUArchitecture.isAppleSilicon ? .metal : .cpu
+ }
+
+ private func unloadModel() {
+ self.stateLock.lock()
+ defer { self.stateLock.unlock() }
+ self.session = nil
+ self.model = nil
+ self.ready = false
+ self.loadedModelName = nil
+ }
+
+ private func currentLoadedModelName() -> String? {
+ self.stateLock.lock()
+ defer { self.stateLock.unlock() }
+ return self.loadedModelName
+ }
+
+ private func installModel(_ model: Model, session: Session, modelName: String) {
+ self.stateLock.lock()
+ defer { self.stateLock.unlock() }
+ self.model = model
+ self.session = session
+ self.loadedModelName = modelName
+ self.ready = true
+ }
+
+ private func activeSession() -> Session? {
+ self.stateLock.lock()
+ defer { self.stateLock.unlock() }
+ return self.session
}
- private func isModelFileValid(at url: URL) -> Bool {
+ private func removeLegacyModelIfNeeded() {
+ for legacyFile in SettingsStore.SpeechModel.legacyWhisperModelFiles {
+ let url = self.modelDirectory.appendingPathComponent(legacyFile)
+ guard FileManager.default.fileExists(atPath: url.path) else { continue }
+ do {
+ try FileManager.default.removeItem(at: url)
+ DebugLogger.shared.info("WhisperProvider: Removed legacy Whisper cache \(legacyFile)", source: "WhisperProvider")
+ } catch {
+ DebugLogger.shared.warning(
+ "WhisperProvider: Failed to remove legacy Whisper cache \(legacyFile): \(error.localizedDescription)",
+ source: "WhisperProvider"
+ )
+ }
+ }
+ }
+
+ private func isModelFileValid(at url: URL, for targetModel: SettingsStore.SpeechModel) -> Bool {
+ guard let expectedModelFile = targetModel.whisperModelFile,
+ url.lastPathComponent == expectedModelFile
+ else {
+ return false
+ }
guard FileManager.default.fileExists(atPath: url.path) else { return false }
guard let attributes = try? FileManager.default.attributesOfItem(atPath: url.path),
let size = attributes[.size] as? NSNumber
else {
return false
}
- let sizeBytes = size.int64Value
- guard sizeBytes > 0 else { return false }
- let expectedBytes = SettingsStore.SpeechModel.allCases
- .first { $0.whisperModelFile == url.lastPathComponent }?
- .expectedDownloadBytes
- return expectedBytes.map { sizeBytes == $0 } ?? false
+ return size.int64Value == targetModel.expectedDownloadBytes
}
func prepare(progressHandler: ((ModelPreparationProgress) -> Void)? = nil) async throws {
try Task.checkCancellation()
- // CRITICAL: Capture the target model at start to use consistently throughout this method.
- // This prevents race conditions where SettingsStore could change after await points.
- let targetModel = self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
- let currentModelName = targetModel.whisperModelFile?.trimmingCharacters(in: .whitespacesAndNewlines) ?? "ggml-base.bin"
-
- // Detect model change: if a different model is now selected, force reload
- if self.isReady, self.loadedModelName != currentModelName {
- DebugLogger.shared.info("WhisperProvider: Model changed from \(self.loadedModelName ?? "nil") to \(currentModelName), forcing reload", source: "WhisperProvider")
- self.isReady = false
- self.whisper = nil
- self.loadedModelName = nil
+
+ let targetModel = self.selectedModel
+ let currentModelName = targetModel.whisperModelFile ?? "whisper-base-Q8_0.gguf"
+
+ let loadedModelName = self.currentLoadedModelName()
+ if self.isReady, loadedModelName != currentModelName {
+ DebugLogger.shared.info(
+ "WhisperProvider: Model changed from \(loadedModelName ?? "nil") to \(currentModelName), forcing reload",
+ source: "WhisperProvider"
+ )
+ self.unloadModel()
}
- guard self.isReady == false else { return }
+ guard !self.isReady else { return }
- DebugLogger.shared.info("WhisperProvider: Starting model preparation", source: "WhisperProvider")
+ try Self.backendInitialization.get()
+ try self.validateBackendAvailability(for: targetModel)
- // Ensure model directory exists
try FileManager.default.createDirectory(at: self.modelDirectory, withIntermediateDirectories: true)
if FileManager.default.fileExists(atPath: self.modelURL.path),
- !self.isModelFileValid(at: self.modelURL)
+ !self.isModelFileValid(at: self.modelURL, for: targetModel)
{
DebugLogger.shared.warning(
"WhisperProvider: Found invalid model file at \(self.modelURL.path); removing to force re-download",
@@ -106,28 +179,26 @@ final class WhisperProvider: TranscriptionProvider {
try? FileManager.default.removeItem(at: self.modelURL)
}
- // Download model if not present
if !FileManager.default.fileExists(atPath: self.modelURL.path) {
- DebugLogger.shared.info("WhisperProvider: Downloading Whisper model...", source: "WhisperProvider")
+ DebugLogger.shared.info("WhisperProvider: Downloading Whisper GGUF model...", source: "WhisperProvider")
progressHandler?(.preparingDownload)
try await self.downloadModel { progress in
progressHandler?(.downloading(progress))
}
}
- guard self.isModelFileValid(at: self.modelURL) else {
+ guard self.isModelFileValid(at: self.modelURL, for: targetModel) else {
+ try? FileManager.default.removeItem(at: self.modelURL)
throw NSError(
domain: "WhisperProvider",
code: -3,
userInfo: [NSLocalizedDescriptionKey: "Whisper model file is missing or corrupted. Please re-download the model."]
)
}
+ self.removeLegacyModelIfNeeded()
- // Check available memory before loading large models
- // Use the captured targetModel to ensure consistent memory validation
let requiredMemoryGB = targetModel.requiredMemoryGB
let availableMemoryGB = Self.availableMemoryGB()
-
DebugLogger.shared.info(
"WhisperProvider: Memory check - Required: \(String(format: "%.1f", requiredMemoryGB))GB, Available: \(String(format: "%.1f", availableMemoryGB))GB",
source: "WhisperProvider"
@@ -139,11 +210,9 @@ final class WhisperProvider: TranscriptionProvider {
Required: \(String(format: "%.1f", requiredMemoryGB)) GB
Available: \(String(format: "%.1f", availableMemoryGB)) GB
- Please try a smaller model (e.g., Whisper Base or Small) or close other applications to free up memory.
+ Please try a smaller model or close other applications to free up memory.
"""
-
DebugLogger.shared.error("WhisperProvider: \(errorMessage)", source: "WhisperProvider")
-
throw NSError(
domain: "WhisperProvider",
code: -4,
@@ -151,18 +220,52 @@ final class WhisperProvider: TranscriptionProvider {
)
}
- // Load the model
- DebugLogger.shared.info("WhisperProvider: Loading Whisper model...", source: "WhisperProvider")
+ DebugLogger.shared.info("WhisperProvider: Loading \(currentModelName) with \(self.backend)", source: "WhisperProvider")
progressHandler?(.loading)
- self.whisper = Whisper(fromFileURL: self.modelURL)
+
+ let loadedModel = try Model(
+ path: self.modelURL.path,
+ options: ModelOptions(backend: self.backend)
+ )
+ let runtimeBackend = loadedModel.backend.lowercased()
+ if CPUArchitecture.isAppleSilicon,
+ !runtimeBackend.contains("metal"),
+ !runtimeBackend.contains("mtl")
+ {
+ throw NSError(
+ domain: "WhisperProvider",
+ code: -7,
+ userInfo: [NSLocalizedDescriptionKey: "\(targetModel.displayName) loaded on \(loadedModel.backend), but Metal is required on Apple Silicon."]
+ )
+ }
+ let loadedSession = try loadedModel.session()
try Task.checkCancellation()
- self.loadedModelName = currentModelName
- self.isReady = true
- DebugLogger.shared.info("WhisperProvider: Model ready (\(currentModelName))", source: "WhisperProvider")
+ self.installModel(loadedModel, session: loadedSession, modelName: currentModelName)
+ DebugLogger.shared.info(
+ "WhisperProvider: Model ready (\(currentModelName), backend=\(loadedModel.backend), arch=\(loadedModel.arch))",
+ source: "WhisperProvider"
+ )
+ }
+
+ private func validateBackendAvailability(for model: SettingsStore.SpeechModel) throws {
+ if CPUArchitecture.isAppleSilicon, !Transcribe.backendAvailable(.metal) {
+ throw NSError(
+ domain: "WhisperProvider",
+ code: -5,
+ userInfo: [NSLocalizedDescriptionKey: "\(model.displayName) requires the Metal Whisper backend on Apple Silicon."]
+ )
+ }
+
+ if !CPUArchitecture.isAppleSilicon, !Transcribe.backendAvailable(.cpu) {
+ throw NSError(
+ domain: "WhisperProvider",
+ code: -6,
+ userInfo: [NSLocalizedDescriptionKey: "Whisper CPU backend is unavailable on this Mac."]
+ )
+ }
}
- /// Returns the available system memory in GB
private static func availableMemoryGB() -> Double {
var pageSize: vm_size_t = 0
host_page_size(mach_host_self(), &pageSize)
@@ -177,24 +280,18 @@ final class WhisperProvider: TranscriptionProvider {
}
guard result == KERN_SUCCESS else {
- // Fallback: assume we have enough memory if we can't check
DebugLogger.shared.warning("WhisperProvider: Failed to get memory stats, assuming sufficient memory", source: "WhisperProvider")
return 16.0
}
- // Calculate free + inactive memory (memory that can be reclaimed)
let freePages = UInt64(vmStats.free_count)
let inactivePages = UInt64(vmStats.inactive_count)
let purgablePages = UInt64(vmStats.purgeable_count)
-
let availableBytes = (freePages + inactivePages + purgablePages) * UInt64(pageSize)
- let availableGB = Double(availableBytes) / (1024 * 1024 * 1024)
-
- return availableGB
+ return Double(availableBytes) / (1024 * 1024 * 1024)
}
func transcribe(_ samples: [Float]) async throws -> ASRTranscriptionResult {
- // Whisper.cpp asserts on very short buffers; guard early to avoid abort.
let minSamples = 16_000
guard samples.count >= minSamples else {
throw NSError(
@@ -204,7 +301,7 @@ final class WhisperProvider: TranscriptionProvider {
)
}
- guard let whisper = whisper else {
+ guard let session = self.activeSession() else {
throw NSError(
domain: "WhisperProvider",
code: -1,
@@ -212,50 +309,48 @@ final class WhisperProvider: TranscriptionProvider {
)
}
- // SwiftWhisper expects 16kHz PCM audio frames (which is what we receive)
- let segments = try await whisper.transcribe(audioFrames: samples)
-
- // Combine all segments into one string
- let fullText = segments.map { $0.text }.joined(separator: " ").trimmingCharacters(in: CharacterSet.whitespacesAndNewlines)
-
- // SwiftWhisper doesn't provide confidence, so we use 1.0
+ let transcript = try await session.run(
+ samples,
+ options: RunOptions(timestamps: .segment)
+ )
+ let fullText = transcript.text.trimmingCharacters(in: .whitespacesAndNewlines)
return ASRTranscriptionResult(text: fullText, confidence: 1.0)
}
func modelsExistOnDisk() -> Bool {
- return self.isModelFileValid(at: self.modelURL)
+ try? FileManager.default.createDirectory(at: self.modelDirectory, withIntermediateDirectories: true)
+ return self.isModelFileValid(at: self.modelURL, for: self.selectedModel)
}
func clearCache() async throws {
+ self.unloadModel()
+
if FileManager.default.fileExists(atPath: self.modelURL.path) {
try FileManager.default.removeItem(at: self.modelURL)
}
+ if let legacyModelURL, FileManager.default.fileExists(atPath: legacyModelURL.path) {
+ try FileManager.default.removeItem(at: legacyModelURL)
+ }
+ self.removeLegacyModelIfNeeded()
+
if FileManager.default.fileExists(atPath: self.modelDirectory.path) {
let contents = try FileManager.default.contentsOfDirectory(atPath: self.modelDirectory.path)
if contents.isEmpty {
try FileManager.default.removeItem(at: self.modelDirectory)
}
}
- self.isReady = false
- self.whisper = nil
- self.loadedModelName = nil
}
- // MARK: - Model Download
-
private func downloadModel(progressHandler: ((Double) -> Void)?) async throws {
- // Whisper models are hosted on Hugging Face
- let modelURLString = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/\(modelName)"
-
- guard let url = URL(string: modelURLString) else {
+ guard let url = self.modelDownloadURL else {
throw NSError(
domain: "WhisperProvider",
code: -1,
- userInfo: [NSLocalizedDescriptionKey: "Invalid model URL"]
+ userInfo: [NSLocalizedDescriptionKey: "Invalid Whisper model URL"]
)
}
- DebugLogger.shared.info("WhisperProvider: Downloading from \(modelURLString)", source: "WhisperProvider")
+ DebugLogger.shared.info("WhisperProvider: Downloading from \(url.absoluteString)", source: "WhisperProvider")
let maxAttempts = 3
for attempt in 1...maxAttempts {
@@ -263,9 +358,7 @@ final class WhisperProvider: TranscriptionProvider {
if attempt == 1 {
progressHandler?(0.0)
}
-
try await self.downloadFile(from: url, to: self.modelURL, progressHandler: progressHandler)
-
DebugLogger.shared.info("WhisperProvider: Model downloaded successfully", source: "WhisperProvider")
return
} catch let error as NSError {
@@ -274,9 +367,8 @@ final class WhisperProvider: TranscriptionProvider {
{
throw CancellationError()
}
- let isLastAttempt = attempt == maxAttempts
- // Provide user-friendly error messages
+ let isLastAttempt = attempt == maxAttempts
if error.domain == NSURLErrorDomain {
let message: String
switch error.code {
@@ -297,7 +389,6 @@ final class WhisperProvider: TranscriptionProvider {
userInfo: [NSLocalizedDescriptionKey: message]
)
}
-
DebugLogger.shared.warning(
"WhisperProvider: Download attempt \(attempt)/\(maxAttempts) failed (\(message)). Retrying...",
source: "WhisperProvider"
@@ -310,7 +401,6 @@ final class WhisperProvider: TranscriptionProvider {
)
}
- // Backoff: 1s, 2s, 4s
let delayNanos = UInt64(1_000_000_000) << UInt64(attempt - 1)
try await Task.sleep(nanoseconds: delayNanos)
}
@@ -349,25 +439,20 @@ final class WhisperProvider: TranscriptionProvider {
let attributes = try FileManager.default.attributesOfItem(atPath: downloadedURL.path)
let actualBytes = (attributes[.size] as? NSNumber)?.int64Value ?? 0
- if httpResponse.expectedContentLength > 0,
- actualBytes != httpResponse.expectedContentLength
- {
+ guard actualBytes > 0 else {
throw NSError(
domain: "WhisperProvider",
code: -1,
- userInfo: [NSLocalizedDescriptionKey: "Downloaded model size mismatch. Please try again."]
+ userInfo: [NSLocalizedDescriptionKey: "Downloaded model is empty. Please try again."]
)
}
- guard
- let expectedBytes = SettingsStore.SpeechModel.allCases
- .first(where: { $0.whisperModelFile == destination.lastPathComponent })?
- .expectedDownloadBytes,
- actualBytes == expectedBytes
- else {
+ if httpResponse.expectedContentLength > 0,
+ actualBytes != httpResponse.expectedContentLength
+ {
throw NSError(
domain: "WhisperProvider",
code: -1,
- userInfo: [NSLocalizedDescriptionKey: "Downloaded model is invalid. Please try again."]
+ userInfo: [NSLocalizedDescriptionKey: "Downloaded model size mismatch. Please try again."]
)
}
diff --git a/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift b/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
index 6c38f3bd..4f5761fe 100644
--- a/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
+++ b/Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
@@ -874,6 +874,38 @@ final class DictationE2ETests: XCTestCase {
)
}
+ func testWhisperProvider_legacyBinCacheDoesNotCountAsDownloadedOrDeletedByReadinessCheck() throws {
+ let modelDirectory = Self.modelDirectoryForRun()
+ try FileManager.default.createDirectory(at: modelDirectory, withIntermediateDirectories: true)
+
+ let legacyURL = modelDirectory.appendingPathComponent("ggml-tiny.bin")
+ try Data([0x01, 0x02, 0x03]).write(to: legacyURL)
+
+ let provider = WhisperProvider(modelDirectory: modelDirectory, modelOverride: .whisperTiny)
+
+ XCTAssertFalse(provider.modelsExistOnDisk())
+ XCTAssertTrue(FileManager.default.fileExists(atPath: legacyURL.path))
+ }
+
+ func testWhisperProvider_ggufCacheReadinessDoesNotDeleteLegacyUntilExplicitClear() async throws {
+ let modelDirectory = Self.modelDirectoryForRun()
+ try FileManager.default.createDirectory(at: modelDirectory, withIntermediateDirectories: true)
+
+ let model = SettingsStore.SpeechModel.whisperTiny
+ let ggufURL = modelDirectory.appendingPathComponent(model.whisperModelFile!)
+ let legacyURL = modelDirectory.appendingPathComponent(model.legacyWhisperModelFile!)
+ try Self.createSparseFile(at: ggufURL, size: model.expectedDownloadBytes)
+ try Data([0x01, 0x02, 0x03]).write(to: legacyURL)
+
+ let provider = WhisperProvider(modelDirectory: modelDirectory, modelOverride: model)
+
+ XCTAssertTrue(provider.modelsExistOnDisk())
+ XCTAssertTrue(FileManager.default.fileExists(atPath: legacyURL.path))
+ try await provider.clearCache()
+ XCTAssertFalse(FileManager.default.fileExists(atPath: ggufURL.path))
+ XCTAssertFalse(FileManager.default.fileExists(atPath: legacyURL.path))
+ }
+
func testAppPromptBinding_profileOverridesModeSelection() {
self.withPromptSettingsRestored {
let settings = SettingsStore.shared
@@ -1520,6 +1552,13 @@ final class DictationE2ETests: XCTestCase {
return base.appendingPathComponent("WhisperModels", isDirectory: true)
}
+ private static func createSparseFile(at url: URL, size: Int64) throws {
+ _ = FileManager.default.createFile(atPath: url.path, contents: nil)
+ let handle = try FileHandle(forWritingTo: url)
+ try handle.truncate(atOffset: UInt64(size))
+ try handle.close()
+ }
+
private static func normalize(_ text: String) -> String {
let lowered = text.lowercased()
let noPunct = lowered.unicodeScalars.map { scalar -> Character in