Fix Cohere long-form transcription

Fluid.xcodeproj/project.pbxproj

@@ -19,6 +19,7 @@
 		272BFB5CB271489892CAE50C /* TemperatureSupportTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 980330F3CE464336ADCE3E23 /* TemperatureSupportTests.swift */; };
 		A62300000000000000000002 /* AudioBufferConverterTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = A62300000000000000000001 /* AudioBufferConverterTests.swift */; };
 		C0DE63600000000000000002 /* AudioEngineRetirementDrainTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */; };
+		C0A3E0010000000000000002 /* TranscribeCppLongFormProcessorTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = C0A3E0010000000000000001 /* TranscribeCppLongFormProcessorTests.swift */; };
 		DA7100020000000000000002 /* DirectAudioReliabilityTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = DA7100010000000000000001 /* DirectAudioReliabilityTests.swift */; };
 		7CDB0A2F2F3C4D5600FB7CAD /* dictation_fixture.wav in Resources */ = {isa = PBXBuildFile; fileRef = 7CDB0A2B2F3C4D5600FB7CAD /* dictation_fixture.wav */; };
 		7CDB0A302F3C4D5600FB7CAD /* XCTest.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 7CDB0A2C2F3C4D5600FB7CAD /* XCTest.framework */; };
@@ -55,6 +56,7 @@
 		980330F3CE464336ADCE3E23 /* TemperatureSupportTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TemperatureSupportTests.swift; sourceTree = "<group>"; };
 		A62300000000000000000001 /* AudioBufferConverterTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AudioBufferConverterTests.swift; sourceTree = "<group>"; };
 		C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AudioEngineRetirementDrainTests.swift; sourceTree = "<group>"; };
+		C0A3E0010000000000000001 /* TranscribeCppLongFormProcessorTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TranscribeCppLongFormProcessorTests.swift; sourceTree = "<group>"; };
 		DA7100010000000000000001 /* DirectAudioReliabilityTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DirectAudioReliabilityTests.swift; sourceTree = "<group>"; };
 		7C078D8F2E3B339200FB7CAC /* FluidVoice Debug.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "FluidVoice Debug.app"; sourceTree = BUILT_PRODUCTS_DIR; };
 		7C91B0022F42AA0100C0DEF0 /* HotkeyShortcutTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = HotkeyShortcutTests.swift; sourceTree = "<group>"; };
@@ -134,6 +136,7 @@
 				980330F3CE464336ADCE3E23 /* TemperatureSupportTests.swift */,
 				A62300000000000000000001 /* AudioBufferConverterTests.swift */,
 				C0DE63600000000000000001 /* AudioEngineRetirementDrainTests.swift */,
+				C0A3E0010000000000000001 /* TranscribeCppLongFormProcessorTests.swift */,
 				DA7100010000000000000001 /* DirectAudioReliabilityTests.swift */,
 			);
 			path = FluidDictationIntegrationTests;
@@ -294,6 +297,7 @@
 				272BFB5CB271489892CAE50C /* TemperatureSupportTests.swift in Sources */,
 				A62300000000000000000002 /* AudioBufferConverterTests.swift in Sources */,
 				C0DE63600000000000000002 /* AudioEngineRetirementDrainTests.swift in Sources */,
+				C0A3E0010000000000000002 /* TranscribeCppLongFormProcessorTests.swift in Sources */,
 				DA7100020000000000000002 /* DirectAudioReliabilityTests.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;

Sources/Fluid/Persistence/SettingsStore.swift

@@ -4205,7 +4205,7 @@ final class SettingsStore: ObservableObject {
             case .parakeetTDTv2: return "~442.9 MiB"
             case .parakeetRealtime: return "~428.4 MiB"
             case .qwen3Asr: return "~2.0 GiB"
-            case .cohereTranscribeSixBit: return "~1.54 GiB"
+            case .cohereTranscribeSixBit: return "~1.45 GiB"
             case .nemotronOffline: return "~530.8 MiB"
             case .nemotronStreaming: return "~668.2 MiB"
             case .nemotronStreaming320: return "~668.2 MiB"
@@ -4226,7 +4226,7 @@ final class SettingsStore: ObservableObject {
             case .parakeetTDTv2: return 464_421_712
             case .parakeetRealtime: return 449_190_189
             case .qwen3Asr: return 2000 * 1024 * 1024
-            case .cohereTranscribeSixBit: return 1_650_748_785
+            case .cohereTranscribeSixBit: return 1_558_162_944
             case .nemotronOffline: return 556_552_620
             case .nemotronStreaming, .nemotronStreaming320: return 700_685_415
             case .whisperTiny: return 45_981_088
@@ -4266,6 +4266,16 @@ final class SettingsStore: ObservableObject {
             }
         }
 
+        /// The GGUF filename for any model served by transcribe.cpp.
+        var transcribeCppModelFile: String? {
+            switch self {
+            case .cohereTranscribeSixBit:
+                return "cohere-transcribe-03-2026-Q4_K_M.gguf"
+            default:
+                return self.whisperModelFile
+            }
+        }
+
         var legacyWhisperModelFile: String? {
             switch self {
             case .whisperTiny: return "ggml-tiny.bin"
@@ -4692,13 +4702,16 @@ final class SettingsStore: ObservableObject {
                 return false
                 #endif
             case .cohereTranscribeSixBit:
-                guard
-                    let spec = self.externalCoreMLSpec,
-                    let directory = SettingsStore.shared.externalCoreMLArtifactsDirectory(for: self)
-                else {
-                    return false
-                }
-                return spec.validatesInstalledArtifacts(at: directory)
+                guard let cacheDirectory = FileManager.default.urls(for: .cachesDirectory, in: .userDomainMask).first,
+                      let modelFile = self.transcribeCppModelFile
+                else { return false }
+                let modelURL = cacheDirectory
+                    .appendingPathComponent("CohereTranscribeModels", isDirectory: true)
+                    .appendingPathComponent(modelFile, isDirectory: false)
+                guard let attributes = try? FileManager.default.attributesOfItem(atPath: modelURL.path),
+                      let size = attributes[.size] as? NSNumber
+                else { return false }
+                return size.int64Value == self.expectedDownloadBytes
             case .nemotronOffline, .nemotronStreaming, .nemotronStreaming320:
                 let hint: String
                 switch self {

Sources/Fluid/Services/ASRService.swift

@@ -172,6 +172,7 @@ final class ASRService: ObservableObject {
     }
 
     @Published var isRunning: Bool = false
+    @Published private(set) var isFileTranscriptionRunning: Bool = false
     @Published var finalText: String = ""
     @Published var partialTranscription: String = ""
     @Published var wordBoostStatusText: String = "Word boost: off"
@@ -195,6 +196,20 @@ final class ASRService: ObservableObject {
         self.isRunning || self.isStarting
     }
 
+    var isAnyTranscriptionRunning: Bool {
+        self.isRunningOrStarting || self.isFileTranscriptionRunning
+    }
+
+    func beginFileTranscription() -> Bool {
+        guard !self.isAnyTranscriptionRunning else { return false }
+        self.isFileTranscriptionRunning = true
+        return true
+    }
+
+    func endFileTranscription() {
+        self.isFileTranscriptionRunning = false
+    }
+
     private let audioCaptureReadinessGate = AudioCaptureReadinessGate()
     private let firstPCMTimeoutNanoseconds: UInt64 = 2_000_000_000
     private var audioCaptureStartGeneration: UInt64 = 0
@@ -353,7 +368,7 @@ final class ASRService: ObservableObject {
         case .parakeetRealtime:
             return self.getParakeetRealtimeProvider()
         case .cohereTranscribeSixBit:
-            return self.getExternalCoreMLProvider()
+            return self.getWhisperProvider()
         case .nemotronOffline, .nemotronStreaming, .nemotronStreaming320:
             return self.getNemotronProvider(mode: model.nemotronProviderMode)
         case .qwen3Asr:
@@ -554,7 +569,7 @@ final class ASRService: ObservableObject {
         case .parakeetRealtime:
             return ParakeetRealtimeProvider()
         case .cohereTranscribeSixBit:
-            return ExternalCoreMLTranscriptionProvider(modelOverride: model)
+            return WhisperProvider(modelOverride: model)
         case .nemotronOffline, .nemotronStreaming, .nemotronStreaming320:
             return NemotronProvider(mode: model.nemotronProviderMode)
         case .qwen3Asr:
@@ -571,6 +586,9 @@ 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 {
+        guard !self.isFileTranscriptionRunning else {
+            throw self.fileTranscriptionModelOperationError()
+        }
         guard self.modelDownloadTask == nil, self.ensureReadyTask == nil else {
             throw NSError(
                 domain: "ASRService",
@@ -644,6 +662,13 @@ final class ASRService: ObservableObject {
 
     /// Call this when the transcription provider setting changes to reset state
     func resetTranscriptionProvider() {
+        guard !self.isFileTranscriptionRunning else {
+            DebugLogger.shared.warning(
+                "ASRService: Ignoring model reset during file transcription",
+                source: "ASRService"
+            )
+            return
+        }
         let newModel = SettingsStore.shared.selectedSpeechModel
         DebugLogger.shared.info("ASRService: Switching to '\(newModel.displayName)', resetting provider state...", source: "ASRService")
 
@@ -1386,6 +1411,10 @@ final class ASRService: ObservableObject {
     ) async -> AudioCaptureStartOutcome {
         DebugLogger.shared.info("🎤 START() called - beginning recording session", source: "ASRService")
 
+        guard !self.isFileTranscriptionRunning else {
+            DebugLogger.shared.warning("START() blocked - file transcription is active", source: "ASRService")
+            return .failed
+        }
         guard self.micStatus == .authorized else {
             DebugLogger.shared.error("❌ START() blocked - mic not authorized", source: "ASRService")
             return .failed
@@ -3738,6 +3767,9 @@ final class ASRService: ObservableObject {
     // MARK: - Cache management
 
     func clearModelCache() async throws {
+        guard !self.isFileTranscriptionRunning else {
+            throw self.fileTranscriptionModelOperationError()
+        }
         DebugLogger.shared.debug("Clearing model cache via transcription provider", source: "ASRService")
         await self.transcriptionExecutor.cancelAndAwaitPending()
         try await self.transcriptionProvider.clearCache()
@@ -3746,6 +3778,9 @@ final class ASRService: ObservableObject {
     }
 
     func clearModelCache(for model: SettingsStore.SpeechModel) async throws {
+        guard !self.isFileTranscriptionRunning else {
+            throw self.fileTranscriptionModelOperationError()
+        }
         DebugLogger.shared.debug("Clearing model cache for \(model.displayName)", source: "ASRService")
         if SettingsStore.shared.selectedSpeechModel == model {
             await self.transcriptionExecutor.cancelAndAwaitPending()
@@ -3762,6 +3797,14 @@ final class ASRService: ObservableObject {
         await self.checkIfModelsExistAsync()
     }
 
+    private func fileTranscriptionModelOperationError() -> NSError {
+        NSError(
+            domain: "ASRService",
+            code: -2002,
+            userInfo: [NSLocalizedDescriptionKey: "Voice model controls are unavailable during meeting transcription."]
+        )
+    }
+
     // MARK: - Timer-based Streaming Transcription (No VAD)
 
     private func startStreamingTranscription() {

Sources/Fluid/Services/ExternalCoreMLModelRegistry.swift

@@ -206,30 +206,8 @@ struct ExternalCoreMLASRModelSpec {
 }
 
 enum ExternalCoreMLModelRegistry {
-    static func spec(for model: SettingsStore.SpeechModel) -> ExternalCoreMLASRModelSpec? {
-        switch model {
-        case .cohereTranscribeSixBit:
-            return ExternalCoreMLASRModelSpec(
-                backend: .cohereTranscribe,
-                artifactFolderHint: "cohere-transcribe-03-2026-CoreML-6bit",
-                manifestFileName: "coreml_manifest.json",
-                frontendFileName: "cohere_frontend.mlpackage",
-                encoderFileName: "cohere_encoder.mlpackage",
-                crossKVProjectorFileName: "cohere_cross_kv_projector.mlpackage",
-                decoderFileName: "cohere_decoder_fullseq_masked.mlpackage",
-                cachedDecoderFileName: "cohere_decoder_cached.mlpackage",
-                expectedModelID: "CohereLabs/cohere-transcribe-03-2026",
-                expectedSampleRate: 16_000,
-                computeConfiguration: .aneSmall,
-                sourceURL: URL(string: "https://huggingface.co/BarathwajAnandan/cohere-transcribe-03-2026-CoreML-6bit"),
-                repositoryOwner: "BarathwajAnandan",
-                repositoryName: "cohere-transcribe-03-2026-CoreML-6bit",
-                repositoryRevision: "main",
-                artifactBundleVersion: "2026-04-02-cohere-refresh-1"
-            )
-        default:
-            return nil
-        }
+    static func spec(for _: SettingsStore.SpeechModel) -> ExternalCoreMLASRModelSpec? {
+        nil
     }
 }
 

Sources/Fluid/Services/MeetingTranscriptionService.swift

@@ -159,14 +159,19 @@ final class MeetingTranscriptionService: ObservableObject {
     /// - Parameters:
     ///   - fileURL: URL to the audio/video file
     func transcribeFile(_ fileURL: URL) async throws -> TranscriptionResult {
+        guard self.asrService.beginFileTranscription() else {
+            throw TranscriptionError.transcriptionFailed("Another transcription is already in progress.")
+        }
         self.isTranscribing = true
         error = nil
         self.progress = 0.0
         let startTime = Date()
+        let transcriptionModel = SettingsStore.shared.selectedSpeechModel
 
         defer {
             isTranscribing = false
             progress = 0.0
+            self.asrService.endFileTranscription()
         }
 
         do {
@@ -279,6 +284,10 @@ final class MeetingTranscriptionService: ObservableObject {
 
             // Calculate chunk size in source file frames
             let sourceFramesPerChunk = AVAudioFrameCount(Double(samplesPerChunk) / resampleRatio)
+            let usesCohereSeamMerging = transcriptionModel == .cohereTranscribeSixBit
+            let sourceOverlapFrames = usesCohereSeamMerging
+                ? AVAudioFrameCount(2 * fileSampleRate)
+                : 0
             var currentFrame: AVAudioFramePosition = 0
 
             self.currentStatus = duration > 0 ? "Transcribing audio (\(Int(duration))s)..." : "Transcribing audio..."
@@ -325,7 +334,10 @@ final class MeetingTranscriptionService: ObservableObject {
                     chunkCount += 1
                 }
 
-                currentFrame += AVAudioFramePosition(framesToRead)
+                let overlap = currentFrame + AVAudioFramePosition(framesToRead) < audioFile.length
+                    ? min(sourceOverlapFrames, framesToRead - 1)
+                    : 0
+                currentFrame += AVAudioFramePosition(framesToRead - overlap)
 
                 // Update progress
                 let progressPercent = Double(currentFrame) / Double(audioFile.length)
@@ -341,7 +353,9 @@ final class MeetingTranscriptionService: ObservableObject {
             }
 
             // Combine all chunk transcriptions
-            let finalText = allTranscriptions.joined(separator: " ")
+            let finalText = usesCohereSeamMerging
+                ? CohereTranscribeCppLongFormProcessor.merge(allTranscriptions)
+                : allTranscriptions.joined(separator: " ")
             let avgConfidence = chunkCount > 0 ? totalConfidence / Float(chunkCount) : 0
 
             let transcriptionResult = (text: finalText, confidence: avgConfidence)