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: Add memory safety checks and UX improvements for large Whisper models

altic-devfca453b246ee
diff --git a/Sources/Fluid/ContentView.swift b/Sources/Fluid/ContentView.swift
index a6f73ed3..efa94b82 100644
--- a/Sources/Fluid/ContentView.swift
+++ b/Sources/Fluid/ContentView.swift
@@ -1409,13 +1409,24 @@ struct ContentView: View {
// Don't reset overlay mode here - let it stay colored until it hides
}
+ // Show "Transcribing..." state before calling stop() to keep overlay visible.
+ // The asr.stop() call performs the final transcription which can take a moment
+ // (especially for slower models like Whisper Medium/Large).
+ DebugLogger.shared.debug("Showing transcription processing state", source: "ContentView")
+ self.menuBarManager.setProcessing(true)
+ NotchOverlayManager.shared.updateTranscriptionText("Transcribing...")
+
// Stop the ASR service and wait for transcription to complete
- // This will set isRunning = false, which triggers overlay hide
- // The overlay will hide while still showing the correct mode color (no gray transition)
+ // The processing indicator will stay visible during this phase
let transcribedText = await asr.stop()
+ // Reset the transcription text display after transcription completes
+ NotchOverlayManager.shared.updateTranscriptionText("")
+
guard transcribedText.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty == false else {
DebugLogger.shared.debug("Transcription returned empty text", source: "ContentView")
+ // Hide processing state when returning early
+ self.menuBarManager.setProcessing(false)
return
}
@@ -1428,11 +1439,12 @@ struct ContentView: View {
guard DictationAIPostProcessingGate.isConfigured() else {
promptTest.lastError = "AI post-processing is not configured. Enable AI Enhancement and configure a provider/model (and API key for non-local endpoints)."
+ self.menuBarManager.setProcessing(false)
return
}
promptTest.isProcessing = true
- self.menuBarManager.setProcessing(true)
+ // Processing already true from above
defer {
self.menuBarManager.setProcessing(false)
promptTest.isProcessing = false
@@ -1482,8 +1494,8 @@ struct ContentView: View {
if shouldUseAI {
DebugLogger.shared.debug("Routing transcription through AI post-processing", source: "ContentView")
- // Show processing animation in notch
- self.menuBarManager.setProcessing(true)
+ // Update overlay text to show we're now refining (processing already true)
+ NotchOverlayManager.shared.updateTranscriptionText("Refining...")
finalText = await self.processTextWithAI(transcribedText)
@@ -1491,6 +1503,8 @@ struct ContentView: View {
self.menuBarManager.setProcessing(false)
} else {
finalText = transcribedText
+ // No AI processing, hide the processing state
+ self.menuBarManager.setProcessing(false)
}
// Apply GAAV formatting as the FINAL step (after AI post-processing)
diff --git a/Sources/Fluid/Persistence/SettingsStore.swift b/Sources/Fluid/Persistence/SettingsStore.swift
index a3f6494c..53ab1c94 100644
--- a/Sources/Fluid/Persistence/SettingsStore.swift
+++ b/Sources/Fluid/Persistence/SettingsStore.swift
@@ -1453,7 +1453,7 @@ final class SettingsStore: ObservableObject {
case whisperBase = "whisper-base"
case whisperSmall = "whisper-small"
case whisperMedium = "whisper-medium"
- // case whisperLargeTurbo = "whisper-large-turbo" // buggy - so removed temporarily
+ case whisperLargeTurbo = "whisper-large-turbo" // temporarily disabled in UI
case whisperLarge = "whisper-large"
var id: String { rawValue }
@@ -1470,7 +1470,7 @@ 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" // buggy - so removed temporarily
+ case .whisperLargeTurbo: return "Whisper Large Turbo (Disabled)"
case .whisperLarge: return "Whisper Large"
}
}
@@ -1481,7 +1481,7 @@ final class SettingsStore: ObservableObject {
case .parakeetTDTv2: return "English Only (Higher Accuracy)"
case .appleSpeech: return "System Languages"
case .appleSpeechAnalyzer: return "EN, ES, FR, DE, IT, JA, KO, PT, ZH"
- case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, /* .whisperLargeTurbo, */ .whisperLarge: // buggy - so removed temporarily
+ case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, .whisperLargeTurbo, .whisperLarge:
return "99 Languages"
}
}
@@ -1496,7 +1496,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return "~142 MB"
case .whisperSmall: return "~466 MB"
case .whisperMedium: return "~1.5 GB"
- // case .whisperLargeTurbo: return "~1.6 GB" // buggy - so removed temporarily
+ case .whisperLargeTurbo: return "~1.6 GB"
case .whisperLarge: return "~2.9 GB"
}
}
@@ -1522,7 +1522,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return "ggml-base.bin"
case .whisperSmall: return "ggml-small.bin"
case .whisperMedium: return "ggml-medium.bin"
- // case .whisperLargeTurbo: return "ggml-large-v3-turbo.bin" // buggy - so removed temporarily
+ case .whisperLargeTurbo: return "ggml-large-v3-turbo.bin"
case .whisperLarge: return "ggml-large-v3.bin"
default: return nil
}
@@ -1535,7 +1535,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return "base"
case .whisperSmall: return "small"
case .whisperMedium: return "medium"
- // case .whisperLargeTurbo: return "large-v3-turbo" // buggy - so removed temporarily
+ case .whisperLargeTurbo: return "large-v3-turbo"
case .whisperLarge: return "large-v3"
default: return nil
}
@@ -1554,6 +1554,9 @@ 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 {
+ return false
+ }
// Filter by Apple Silicon requirement
if model.requiresAppleSilicon, !CPUArchitecture.isAppleSilicon {
return false
@@ -1588,7 +1591,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return "Standard Choice"
case .whisperSmall: return "Balanced Speed & Accuracy"
case .whisperMedium: return "Medium Quality"
- // case .whisperLargeTurbo: return "Higher Quality but Faster" // buggy - so removed temporarily
+ case .whisperLargeTurbo: return "Higher Quality but Faster"
case .whisperLarge: return "Maximum Accuracy"
}
}
@@ -1612,13 +1615,49 @@ final class SettingsStore: ObservableObject {
return "Better accuracy than Base. Moderate resource usage."
case .whisperMedium:
return "High accuracy for demanding tasks. Requires more memory."
- // case .whisperLargeTurbo: // buggy - so removed temporarily
- // return "Near-maximum accuracy with optimized speed."
+ case .whisperLargeTurbo:
+ return "Near-maximum accuracy with optimized speed."
case .whisperLarge:
return "Best possible accuracy. Large download and memory usage."
}
}
+ /// Minimum recommended RAM in GB for this model to run safely
+ var requiredMemoryGB: Double {
+ switch self {
+ case .parakeetTDT, .parakeetTDTv2:
+ return 4.0
+ case .appleSpeech, .appleSpeechAnalyzer:
+ return 2.0 // Built-in, minimal overhead
+ case .whisperTiny:
+ return 2.0
+ case .whisperBase:
+ return 3.0
+ case .whisperSmall:
+ return 4.0
+ case .whisperMedium:
+ return 6.0
+ case .whisperLargeTurbo:
+ return 8.0
+ case .whisperLarge:
+ return 10.0 // Large model needs ~6-8GB working memory + model size
+ }
+ }
+
+ /// Warning text for models with high memory requirements, nil if no warning needed
+ var memoryWarning: String? {
+ switch self {
+ case .whisperLarge:
+ return "⚠️ Requires 10GB+ RAM. May crash on systems with limited memory."
+ case .whisperLargeTurbo:
+ return "⚠️ Requires 8GB+ RAM. May be unstable on some systems."
+ case .whisperMedium:
+ return "Requires 6GB+ RAM for stable operation."
+ default:
+ return nil
+ }
+ }
+
/// Speed rating (1-5, higher is faster)
var speedRating: Int {
switch self {
@@ -1630,7 +1669,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return 4
case .whisperSmall: return 3
case .whisperMedium: return 2
- // case .whisperLargeTurbo: return 3 // buggy - so removed temporarily
+ case .whisperLargeTurbo: return 3
case .whisperLarge: return 1
}
}
@@ -1646,7 +1685,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return 3
case .whisperSmall: return 4
case .whisperMedium: return 4
- // case .whisperLargeTurbo: return 5 // buggy - so removed temporarily
+ case .whisperLargeTurbo: return 5
case .whisperLarge: return 5
}
}
@@ -1662,7 +1701,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return 0.80
case .whisperSmall: return 0.60
case .whisperMedium: return 0.40
- // case .whisperLargeTurbo: return 0.65 // buggy - so removed temporarily
+ case .whisperLargeTurbo: return 0.65
case .whisperLarge: return 0.20
}
}
@@ -1678,7 +1717,7 @@ final class SettingsStore: ObservableObject {
case .whisperBase: return 0.60
case .whisperSmall: return 0.70
case .whisperMedium: return 0.80
- // case .whisperLargeTurbo: return 0.95 // buggy - so removed temporarily
+ case .whisperLargeTurbo: return 0.95
case .whisperLarge: return 1.00
}
}
@@ -1707,7 +1746,7 @@ final class SettingsStore: ObservableObject {
/// Large Whisper models are too slow for streaming, so they only do final transcription on stop.
var supportsStreaming: Bool {
switch self {
- case .whisperMedium, .whisperLarge: // .whisperLargeTurbo - buggy - so removed temporarily
+ case .whisperMedium, .whisperLargeTurbo, .whisperLarge:
return false // Too slow for real-time chunk processing
default:
return true // All other models support streaming
@@ -1728,7 +1767,7 @@ final class SettingsStore: ObservableObject {
return .nvidia
case .appleSpeech, .appleSpeechAnalyzer:
return .apple
- case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, /* .whisperLargeTurbo, */ .whisperLarge: // buggy - so removed temporarily
+ case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, .whisperLargeTurbo, .whisperLarge:
return .openai
}
}
@@ -1778,7 +1817,7 @@ final class SettingsStore: ObservableObject {
return "NVIDIA"
case .appleSpeech, .appleSpeechAnalyzer:
return "Apple"
- case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, /* .whisperLargeTurbo, */ .whisperLarge: // buggy - so removed temporarily
+ case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, .whisperLargeTurbo, .whisperLarge:
return "OpenAI"
}
}
@@ -1798,7 +1837,7 @@ final class SettingsStore: ObservableObject {
return "#76B900"
case .appleSpeech, .appleSpeechAnalyzer:
return "#A2AAAD" // Apple Gray
- case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, /* .whisperLargeTurbo, */ .whisperLarge: // buggy - so removed temporarily
+ case .whisperTiny, .whisperBase, .whisperSmall, .whisperMedium, .whisperLargeTurbo, .whisperLarge:
return "#10A37F" // OpenAI Teal
}
}
diff --git a/Sources/Fluid/Services/ASRService.swift b/Sources/Fluid/Services/ASRService.swift
index 7c22bdeb..134166f2 100644
--- a/Sources/Fluid/Services/ASRService.swift
+++ b/Sources/Fluid/Services/ASRService.swift
@@ -40,6 +40,24 @@ private actor TranscriptionExecutor {
}
}
+private actor ModelDownloadRegistry {
+ private var tasks: [String: Task<Void, Error>] = [:]
+
+ func run(for key: String, operation: @escaping () async throws -> Void) async throws {
+ if let existing = tasks[key] {
+ return try await existing.value
+ }
+
+ let task = Task {
+ try await operation()
+ }
+ self.tasks[key] = task
+ defer { tasks[key] = nil }
+
+ try await task.value
+ }
+}
+
/// A comprehensive speech recognition service that handles real-time audio transcription.
///
/// This service manages the entire ASR (Automatic Speech Recognition) pipeline including:
@@ -93,6 +111,7 @@ final class ASRService: ObservableObject {
private var downloadProgressTask: Task<Void, Never>?
private var hasCompletedFirstTranscription: Bool = false // Track if model has warmed up with first transcription
+ private let downloadRegistry = ModelDownloadRegistry()
// MARK: - Error Handling
@@ -225,18 +244,35 @@ final class ASRService: ObservableObject {
/// - model: The model to download
/// - progressHandler: Optional callback for download progress (0.0 to 1.0)
func downloadModel(_ model: SettingsStore.SpeechModel, progressHandler: ((Double) -> Void)?) async throws {
- DebugLogger.shared.info("Downloading model: \(model.displayName) (without changing active selection)", source: "ASRService")
+ try await self.downloadRegistry.run(for: model.id) { [weak self] in
+ guard let self else { return }
- // Get a fresh provider for this specific model (uses modelOverride for Whisper)
- let provider = self.getProvider(for: model)
+ await MainActor.run {
+ self.downloadingModelId = model.id
+ }
- // Prepare (download) the model
- try await provider.prepare(progressHandler: { progress in
- let clamped = max(0.0, min(1.0, progress))
- progressHandler?(clamped)
- })
+ defer {
+ Task { @MainActor [weak self] in
+ guard let self else { return }
+ if self.downloadingModelId == model.id {
+ self.downloadingModelId = nil
+ }
+ }
+ }
+
+ DebugLogger.shared.info("Downloading model: \(model.displayName) (without changing active selection)", source: "ASRService")
- DebugLogger.shared.info("Model download completed: \(model.displayName)", source: "ASRService")
+ // Get a fresh provider for this specific model (uses modelOverride for Whisper)
+ let provider = await MainActor.run { self.getProvider(for: model) }
+
+ // Prepare (download) the model
+ try await provider.prepare(progressHandler: { progress in
+ let clamped = max(0.0, min(1.0, progress))
+ progressHandler?(clamped)
+ })
+
+ DebugLogger.shared.info("Model download completed: \(model.displayName)", source: "ASRService")
+ }
}
/// Call this when the transcription provider setting changes to reset state
diff --git a/Sources/Fluid/Services/FluidAudioProvider.swift b/Sources/Fluid/Services/FluidAudioProvider.swift
index 0822d526..fe125c4c 100644
--- a/Sources/Fluid/Services/FluidAudioProvider.swift
+++ b/Sources/Fluid/Services/FluidAudioProvider.swift
@@ -64,8 +64,9 @@ final class FluidAudioProvider: TranscriptionProvider {
func modelsExistOnDisk() -> Bool {
let baseCacheDir = AsrModels.defaultCacheDirectory().deletingLastPathComponent()
+ let selectedModel = self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
- if SettingsStore.shared.selectedSpeechModel == .parakeetTDTv2 {
+ if selectedModel == .parakeetTDTv2 {
let v2CacheDir = baseCacheDir.appendingPathComponent("parakeet-tdt-0.6b-v2-coreml")
return FileManager.default.fileExists(atPath: v2CacheDir.path)
} else {
@@ -76,7 +77,7 @@ final class FluidAudioProvider: TranscriptionProvider {
func clearCache() async throws {
let baseCacheDir = AsrModels.defaultCacheDirectory().deletingLastPathComponent()
- let selectedModel = SettingsStore.shared.selectedSpeechModel
+ let selectedModel = self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
if selectedModel == .parakeetTDTv2 {
// Clear v2 cache only
diff --git a/Sources/Fluid/Services/GlobalHotkeyManager.swift b/Sources/Fluid/Services/GlobalHotkeyManager.swift
index 7f0e5fc5..24bbf4d2 100644
--- a/Sources/Fluid/Services/GlobalHotkeyManager.swift
+++ b/Sources/Fluid/Services/GlobalHotkeyManager.swift
@@ -3,6 +3,25 @@ import Foundation
@MainActor
final class GlobalHotkeyManager: NSObject {
+ private final class HotkeyState: @unchecked Sendable {
+ private let lock = NSLock()
+ var isKeyPressed = false
+ var isCommandModeKeyPressed = false
+ var isRewriteKeyPressed = false
+ var modifierOnlyKeyDown = false
+ var otherKeyPressedDuringModifier = false
+ var modifierPressStartTime: Date?
+ var pendingHoldModeStart: Task<Void, Never>?
+ var pendingHoldModeType: HoldModeType?
+
+ func withLock<T>(_ block: () -> T) -> T {
+ self.lock.lock()
+ defer { self.lock.unlock() }
+ return block()
+ }
+ }
+
+ private let state = HotkeyState()
private nonisolated(unsafe) var eventTap: CFMachPort?
private nonisolated(unsafe) var runLoopSource: CFRunLoopSource?
private let asrService: ASRService
@@ -17,21 +36,49 @@ final class GlobalHotkeyManager: NSObject {
private var rewriteModeCallback: (() async -> Void)?
private var cancelCallback: (() -> Bool)? // Returns true if handled
private var pressAndHoldMode: Bool = SettingsStore.shared.pressAndHoldMode
- // These state flags are accessed from CGEvent tap callback (may run off main thread)
- // and @MainActor tasks. Marked nonisolated(unsafe) since they're simple booleans
- // with effectively atomic reads/writes, and the event tap runs on main run loop.
- private nonisolated(unsafe) var isKeyPressed = false
- private nonisolated(unsafe) var isCommandModeKeyPressed = false
- private nonisolated(unsafe) var isRewriteKeyPressed = false
+
+ private nonisolated var isKeyPressed: Bool {
+ get { self.state.withLock { self.state.isKeyPressed } }
+ set { self.state.withLock { self.state.isKeyPressed = newValue } }
+ }
+
+ private nonisolated var isCommandModeKeyPressed: Bool {
+ get { self.state.withLock { self.state.isCommandModeKeyPressed } }
+ set { self.state.withLock { self.state.isCommandModeKeyPressed = newValue } }
+ }
+
+ private nonisolated var isRewriteKeyPressed: Bool {
+ get { self.state.withLock { self.state.isRewriteKeyPressed } }
+ set { self.state.withLock { self.state.isRewriteKeyPressed = newValue } }
+ }
// Modifier-only shortcut tracking: detect if another key was pressed during modifier hold
- private nonisolated(unsafe) var modifierOnlyKeyDown = false
- private nonisolated(unsafe) var otherKeyPressedDuringModifier = false
+ private nonisolated var modifierOnlyKeyDown: Bool {
+ get { self.state.withLock { self.state.modifierOnlyKeyDown } }
+ set { self.state.withLock { self.state.modifierOnlyKeyDown = newValue } }
+ }
+
+ private nonisolated var otherKeyPressedDuringModifier: Bool {
+ get { self.state.withLock { self.state.otherKeyPressedDuringModifier } }
+ set { self.state.withLock { self.state.otherKeyPressedDuringModifier = newValue } }
+ }
+
// Reserved for future tap-vs-hold timing detection (e.g., quick tap to toggle vs long hold)
- private var modifierPressStartTime: Date?
- private nonisolated(unsafe) var pendingHoldModeStart: Task<Void, Never>?
+ private nonisolated var modifierPressStartTime: Date? {
+ get { self.state.withLock { self.state.modifierPressStartTime } }
+ set { self.state.withLock { self.state.modifierPressStartTime = newValue } }
+ }
+
+ private nonisolated var pendingHoldModeStart: Task<Void, Never>? {
+ get { self.state.withLock { self.state.pendingHoldModeStart } }
+ set { self.state.withLock { self.state.pendingHoldModeStart = newValue } }
+ }
+
// Tracks which mode's pending start is active (for cancellation on key combos)
- private var pendingHoldModeType: HoldModeType?
+ private nonisolated var pendingHoldModeType: HoldModeType? {
+ get { self.state.withLock { self.state.pendingHoldModeType } }
+ set { self.state.withLock { self.state.pendingHoldModeType = newValue } }
+ }
private enum HoldModeType {
case transcription
diff --git a/Sources/Fluid/Services/MeetingTranscriptionService.swift b/Sources/Fluid/Services/MeetingTranscriptionService.swift
index 4bf864e5..8ec944aa 100644
--- a/Sources/Fluid/Services/MeetingTranscriptionService.swift
+++ b/Sources/Fluid/Services/MeetingTranscriptionService.swift
@@ -218,6 +218,13 @@ final class MeetingTranscriptionService: ObservableObject {
self.currentStatus = "Transcribing... \(Int(progressPercent * 100))%"
}
+ if allTranscriptions.isEmpty {
+ DebugLogger.shared.warning(
+ "No audio chunks were long enough to transcribe (minimum 1 second required)",
+ source: "MeetingTranscriptionService"
+ )
+ }
+
// Combine all chunk transcriptions
let finalText = allTranscriptions.joined(separator: " ")
let avgConfidence = chunkCount > 0 ? totalConfidence / Float(chunkCount) : 0
@@ -377,7 +384,7 @@ final class MeetingTranscriptionService: ObservableObject {
converter.convert(to: outputBuffer, error: &error) { _, outStatus in
if inputConsumed {
- outStatus.pointee = .noDataNow
+ outStatus.pointee = .endOfStream
return nil
}
inputConsumed = true
diff --git a/Sources/Fluid/Services/WhisperProvider.swift b/Sources/Fluid/Services/WhisperProvider.swift
index c22d7c9c..d25825a5 100644
--- a/Sources/Fluid/Services/WhisperProvider.swift
+++ b/Sources/Fluid/Services/WhisperProvider.swift
@@ -136,6 +136,34 @@ final class WhisperProvider: TranscriptionProvider {
)
}
+ // Check available memory before loading large models
+ let selectedModel = self.modelOverride ?? SettingsStore.shared.selectedSpeechModel
+ let requiredMemoryGB = selectedModel.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"
+ )
+
+ if availableMemoryGB < requiredMemoryGB {
+ let errorMessage = """
+ Insufficient memory for \(selectedModel.displayName).
+ 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.
+ """
+
+ DebugLogger.shared.error("WhisperProvider: \(errorMessage)", source: "WhisperProvider")
+
+ throw NSError(
+ domain: "WhisperProvider",
+ code: -4,
+ userInfo: [NSLocalizedDescriptionKey: errorMessage]
+ )
+ }
+
// Load the model
DebugLogger.shared.info("WhisperProvider: Loading Whisper model...", source: "WhisperProvider")
self.whisper = Whisper(fromFileURL: self.modelURL)
@@ -145,6 +173,37 @@ final class WhisperProvider: TranscriptionProvider {
DebugLogger.shared.info("WhisperProvider: Model ready (\(currentModelName))", source: "WhisperProvider")
}
+ /// 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)
+
+ var vmStats = vm_statistics64()
+ var count = mach_msg_type_number_t(MemoryLayout<vm_statistics64>.size / MemoryLayout<integer_t>.size)
+
+ let result = withUnsafeMutablePointer(to: &vmStats) { pointer in
+ pointer.withMemoryRebound(to: integer_t.self, capacity: Int(count)) { intPointer in
+ host_statistics64(mach_host_self(), HOST_VM_INFO64, intPointer, &count)
+ }
+ }
+
+ 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
+ }
+
func transcribe(_ samples: [Float]) async throws -> ASRTranscriptionResult {
// Whisper.cpp asserts on very short buffers; guard early to avoid abort.
let minSamples = 16_000
diff --git a/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift b/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
index e25f5531..00ea45e7 100644
--- a/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
+++ b/Sources/Fluid/UI/AISettings/VoiceEngineSettingsViewModel.swift
@@ -113,6 +113,8 @@ final class VoiceEngineSettingsViewModel: ObservableObject {
func downloadSpeechModel(_ model: SettingsStore.SpeechModel) {
guard !self.asr.isRunning else { return }
guard self.downloadingModel == nil else { return } // Prevent concurrent downloads
+ self.downloadingModel = model
+ self.downloadProgress = 0.0
Task {
// Mark this model as downloading
diff --git a/Sources/Fluid/UI/AISettingsView+SpeechRecognition.swift b/Sources/Fluid/UI/AISettingsView+SpeechRecognition.swift
index 54c410cc..11c2c8a9 100644
--- a/Sources/Fluid/UI/AISettingsView+SpeechRecognition.swift
+++ b/Sources/Fluid/UI/AISettingsView+SpeechRecognition.swift
@@ -198,6 +198,28 @@ extension VoiceEngineSettingsView {
Spacer()
}
+
+ // Memory warning for large models
+ if let memoryWarning = model.memoryWarning {
+ HStack(spacing: 6) {
+ Image(systemName: "exclamationmark.triangle.fill")
+ .font(.caption2)
+ .foregroundStyle(.orange)
+ Text(memoryWarning)
+ .font(.caption2)
+ .foregroundStyle(.orange)
+ }
+ .padding(.horizontal, 8)
+ .padding(.vertical, 4)
+ .background(
+ RoundedRectangle(cornerRadius: 6)
+ .fill(.orange.opacity(0.1))
+ .overlay(
+ RoundedRectangle(cornerRadius: 6)
+ .stroke(.orange.opacity(0.3), lineWidth: 1)
+ )
+ )
+ }
}
.frame(maxWidth: .infinity, alignment: .leading)