fix: preserve system audio timing

Meetingnotes.xcodeproj/project.pbxproj

@@ -276,7 +276,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 33;
+				CURRENT_PROJECT_VERSION = 34;
 				DEVELOPMENT_ASSET_PATHS = "\"meetingnotes/Preview Content\"";
 				DEVELOPMENT_TEAM = G9LVHZAJNX;
 				ENABLE_HARDENED_RUNTIME = YES;
@@ -290,7 +290,7 @@
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 15.0;
-				MARKETING_VERSION = 1.1.21;
+				MARKETING_VERSION = 1.1.22;
 				ONLY_ACTIVE_ARCH = NO;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;
@@ -312,7 +312,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 33;
+				CURRENT_PROJECT_VERSION = 34;
 				DEVELOPMENT_ASSET_PATHS = "\"meetingnotes/Preview Content\"";
 				DEVELOPMENT_TEAM = G9LVHZAJNX;
 				ENABLE_HARDENED_RUNTIME = YES;
@@ -326,7 +326,7 @@
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 15.0;
-				MARKETING_VERSION = 1.1.21;
+				MARKETING_VERSION = 1.1.22;
 				ONLY_ACTIVE_ARCH = YES;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;

meetingnotes/Managers/AudioManager.swift

@@ -86,6 +86,18 @@ final class AudioManager: NSObject, ObservableObject {
             isProcessing = false
         }
 
+        repairHalfDurationSystemWAVIfNeeded(in: files)
+        if let mismatch = captureDurationMismatch(in: files) {
+            let completedFiles = files.compactMap { $0 }
+            let audioFolder = preserveAudioFiles(completedFiles, meetingID: completedMeetingID)
+            lastRecoveryAudioFolderName = audioFolder?.lastPathComponent
+            let recoveryMessage = audioFolder == nil
+                ? " The audio remains in the app's temporary folder."
+                : " Audio was kept so it can be recovered."
+            errorMessage = "System audio timing was invalid (\(mismatch)). Transcription was stopped to avoid an out-of-order result." + recoveryMessage
+            return transcriptChunks.filter(\.isFinal)
+        }
+
         let model = UserDefaultsManager.shared.transcriptionModel
         async let micResult = transcribe(files[0], model: model, diarization: false)
         async let systemResult = transcribe(files[1], model: model, diarization: true)
@@ -356,13 +368,25 @@ final class AudioManager: NSObject, ObservableObject {
     
     private func startTapIO(_ tap: ProcessTap) throws {
         guard var description = tap.tapStreamDescription,
-              let inputFormat = AVAudioFormat(streamDescription: &description),
+              let advertisedInputFormat = AVAudioFormat(streamDescription: &description),
               let targetFormat = systemAudioFile?.processingFormat,
-              let converter = AVAudioConverter(from: inputFormat, to: targetFormat) else {
+              advertisedInputFormat.sampleRate > 0 else {
             throw NSError(domain: "AudioManager", code: -1, userInfo: [NSLocalizedDescriptionKey: "Unsupported system audio format"])
         }
+        var inputFormat: AVAudioFormat?
+        var converter: AVAudioConverter?
         try tap.run(on: tapQueue) { [weak self] _, inputData, _, _, _ in
             guard let self else { return }
+            if inputFormat == nil {
+                inputFormat = self.inputFormat(
+                    for: inputData,
+                    advertisedFormat: advertisedInputFormat
+                )
+                if let inputFormat {
+                    converter = AVAudioConverter(from: inputFormat, to: targetFormat)
+                }
+            }
+            guard let inputFormat, let converter else { return }
             // The tap queue is serial. Reusing the converter preserves its
             // resampler state instead of discarding audio at every callback.
             self.processAudioBuffer(
@@ -377,6 +401,30 @@ final class AudioManager: NSObject, ObservableObject {
         }
     }
 
+    private func inputFormat(
+        for inputData: UnsafePointer<AudioBufferList>,
+        advertisedFormat: AVAudioFormat
+    ) -> AVAudioFormat? {
+        let buffers = UnsafeMutableAudioBufferListPointer(
+            UnsafeMutablePointer(mutating: inputData)
+        )
+        let channelCount = buffers.reduce(UInt32(0)) { $0 + $1.mNumberChannels }
+        guard channelCount > 0 else { return nil }
+
+        // HAL tap metadata can advertise interleaved stereo while the callback
+        // supplies one mono buffer per channel (or the reverse). Constructing a
+        // PCM buffer with that mismatched layout halves its frame count and
+        // produces 2x-speed system audio. The callback's AudioBufferList is the
+        // authoritative layout for the memory we are copying.
+        let isInterleaved = buffers.count == 1 && channelCount > 1
+        return AVAudioFormat(
+            commonFormat: advertisedFormat.commonFormat,
+            sampleRate: advertisedFormat.sampleRate,
+            channels: AVAudioChannelCount(channelCount),
+            interleaved: isInterleaved
+        )
+    }
+
     private func copyAudioBuffer(
         from inputData: UnsafePointer<AudioBufferList>,
         format: AVAudioFormat
@@ -523,6 +571,90 @@ final class AudioManager: NSObject, ObservableObject {
     private func preserveAudioFiles(_ urls: [URL], meetingID: UUID) -> URL? {
         LocalStorageManager.shared.preserveAudioFiles(urls, for: meetingID)
     }
+
+    private func captureDurationMismatch(in files: [URL?]) -> String? {
+        guard files.count >= 2,
+              let micDuration = audioDuration(at: files[0]),
+              let systemDuration = audioDuration(at: files[1]),
+              micDuration >= 60 else { return nil }
+        let ratio = systemDuration / micDuration
+        guard (0.45...0.55).contains(ratio) || (1.8...2.2).contains(ratio) else { return nil }
+        return String(format: "mic %.1fs, system %.1fs", micDuration, systemDuration)
+    }
+
+    private func repairHalfDurationSystemWAVIfNeeded(in files: [URL?]) {
+        guard files.count >= 2,
+              let micDuration = audioDuration(at: files[0]),
+              let systemURL = files[1],
+              let systemDuration = audioDuration(at: systemURL),
+              micDuration >= 60,
+              systemURL.pathExtension.caseInsensitiveCompare("wav") == .orderedSame,
+              (0.48...0.52).contains(systemDuration / micDuration) else { return }
+        try? halveWAVSampleRate(at: systemURL)
+    }
+
+    private func halveWAVSampleRate(at url: URL) throws {
+        let handle = try FileHandle(forUpdating: url)
+        defer { try? handle.close() }
+
+        try handle.seek(toOffset: 0)
+        guard let riffHeader = try handle.read(upToCount: 12),
+              riffHeader.count == 12,
+              String(data: riffHeader[0..<4], encoding: .ascii) == "RIFF",
+              String(data: riffHeader[8..<12], encoding: .ascii) == "WAVE" else {
+            throw NSError(domain: "AudioManager", code: -2, userInfo: [NSLocalizedDescriptionKey: "Invalid WAV header"])
+        }
+
+        var offset: UInt64 = 12
+        while true {
+            try handle.seek(toOffset: offset)
+            guard let chunkHeader = try handle.read(upToCount: 8), chunkHeader.count == 8 else { break }
+            let chunkID = String(data: chunkHeader[0..<4], encoding: .ascii)
+            let chunkSize = UInt32(chunkHeader[4])
+                | (UInt32(chunkHeader[5]) << 8)
+                | (UInt32(chunkHeader[6]) << 16)
+                | (UInt32(chunkHeader[7]) << 24)
+            let chunkDataOffset = offset + 8
+
+            if chunkID == "fmt ", chunkSize >= 16 {
+                try handle.seek(toOffset: chunkDataOffset)
+                guard let format = try handle.read(upToCount: 16), format.count == 16 else { break }
+                let audioFormat = UInt16(format[0]) | (UInt16(format[1]) << 8)
+                let blockAlign = UInt16(format[12]) | (UInt16(format[13]) << 8)
+                let sampleRate = UInt32(format[4])
+                    | (UInt32(format[5]) << 8)
+                    | (UInt32(format[6]) << 16)
+                    | (UInt32(format[7]) << 24)
+                guard audioFormat == 1, sampleRate >= 16_000, sampleRate.isMultiple(of: 2) else {
+                    throw NSError(domain: "AudioManager", code: -3, userInfo: [NSLocalizedDescriptionKey: "Unsupported WAV format"])
+                }
+
+                let correctedSampleRate = sampleRate / 2
+                let correctedByteRate = correctedSampleRate * UInt32(blockAlign)
+                try handle.seek(toOffset: chunkDataOffset + 4)
+                try handle.write(contentsOf: littleEndianData(correctedSampleRate))
+                try handle.write(contentsOf: littleEndianData(correctedByteRate))
+                try handle.synchronize()
+                return
+            }
+
+            offset = chunkDataOffset + UInt64(chunkSize) + UInt64(chunkSize % 2)
+        }
+
+        throw NSError(domain: "AudioManager", code: -4, userInfo: [NSLocalizedDescriptionKey: "WAV format chunk was not found"])
+    }
+
+    private func littleEndianData(_ value: UInt32) -> Data {
+        var littleEndianValue = value.littleEndian
+        return withUnsafeBytes(of: &littleEndianValue) { Data($0) }
+    }
+
+    private func audioDuration(at url: URL?) -> TimeInterval? {
+        guard let url,
+              let file = try? AVAudioFile(forReading: url),
+              file.processingFormat.sampleRate > 0 else { return nil }
+        return Double(file.length) / file.processingFormat.sampleRate
+    }
         
     private func resetAudioLevels() {
         micAudioLevel = 0