fix: count Core Audio callback frames explicitly

Meetingnotes.xcodeproj/project.pbxproj

@@ -276,7 +276,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 36;
+				CURRENT_PROJECT_VERSION = 40;
 				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 = 1.1.24;
+				MARKETING_VERSION = 1.1.28;
 				ONLY_ACTIVE_ARCH = NO;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;
@@ -312,7 +312,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 36;
+				CURRENT_PROJECT_VERSION = 40;
 				DEVELOPMENT_ASSET_PATHS = "\"meetingnotes/Preview Content\"";
 				DEVELOPMENT_TEAM = G9LVHZAJNX;
 				ENABLE_HARDENED_RUNTIME = YES;
@@ -326,7 +326,7 @@
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 15.0;
-				MARKETING_VERSION = 1.1.24;
+				MARKETING_VERSION = 1.1.28;
 				ONLY_ACTIVE_ARCH = YES;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;

README.md

@@ -22,7 +22,7 @@ Implemented:
 - Meeting search functionality
 - Abilty to edit system prompt
 - Select any compatible Coder model for transcription and note generation
-- Automatic start and stop from MuteDeck through a compatible local API
+- Automatic start and stop from MuteDeck through a compatible local API, with a 10-second reconnect grace period
 - Merge an automatically split continuation back into its previous meeting
 - Auto updates
 - Text formatting

meetingnotes/Managers/AudioManager.swift

@@ -443,14 +443,15 @@ final class AudioManager: NSObject, ObservableObject {
         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
+        // HAL I/O proc samples use the canonical Float32 representation. The
-        // supplies one mono buffer per channel (or the reverse). Constructing a
+        // tap's stream description can advertise a different common format;
-        // PCM buffer with that mismatched layout halves its frame count and
+        // using that to interpret the callback bytes can halve the frame count
-        // produces 2x-speed system audio. The callback's AudioBufferList is the
+        // (for example, treating four-byte Float32 samples as eight-byte
-        // authoritative layout for the memory we are copying.
+        // Float64 samples). The callback's AudioBufferList is authoritative for
+        // its channel layout, while its sample rate comes from the input stream.
         let isInterleaved = buffers.count == 1 && channelCount > 1
         return AVAudioFormat(
-            commonFormat: advertisedFormat.commonFormat,
+            commonFormat: .pcmFormatFloat32,
             sampleRate: advertisedFormat.sampleRate,
             channels: AVAudioChannelCount(channelCount),
             interleaved: isInterleaved
@@ -461,20 +462,27 @@ final class AudioManager: NSObject, ObservableObject {
         from inputData: UnsafePointer<AudioBufferList>,
         format: AVAudioFormat
     ) -> AVAudioPCMBuffer? {
-        guard let borrowedBuffer = AVAudioPCMBuffer(
-            pcmFormat: format,
-            bufferListNoCopy: inputData,
-            deallocator: nil
-        ), borrowedBuffer.frameLength > 0,
-        let ownedBuffer = AVAudioPCMBuffer(
-            pcmFormat: format,
-            frameCapacity: borrowedBuffer.frameLength
-        ) else { return nil }
-
-        ownedBuffer.frameLength = borrowedBuffer.frameLength
         let sourceBuffers = UnsafeMutableAudioBufferListPointer(
             UnsafeMutablePointer(mutating: inputData)
         )
+        let frameLengths = sourceBuffers.compactMap { source -> AVAudioFrameCount? in
+            let bytesPerFrame = Int(source.mNumberChannels) * MemoryLayout<Float32>.size
+            guard bytesPerFrame > 0,
+                  Int(source.mDataByteSize).isMultiple(of: bytesPerFrame) else { return nil }
+            return AVAudioFrameCount(Int(source.mDataByteSize) / bytesPerFrame)
+        }
+        guard frameLengths.count == sourceBuffers.count,
+              let frameLength = frameLengths.first,
+              frameLength > 0,
+              frameLengths.allSatisfy({ $0 == frameLength }),
+              let ownedBuffer = AVAudioPCMBuffer(
+                pcmFormat: format,
+                frameCapacity: frameLength
+              ) else { return nil }
+
+        // Set the frame count explicitly instead of asking AVAudioPCMBuffer to
+        // infer it from potentially inconsistent tap metadata.
+        ownedBuffer.frameLength = frameLength
         let destinationBuffers = UnsafeMutableAudioBufferListPointer(
             ownedBuffer.mutableAudioBufferList
         )

meetingnotes/ProcessTap/CoreAudioUtils.swift

@@ -116,6 +116,47 @@ extension AudioObjectID {
         try read(kAudioTapPropertyFormat, defaultValue: AudioStreamBasicDescription())
     }
 
+    /// Reads the virtual format of the first input stream exposed by this device.
+    ///
+    /// An aggregate device can adapt a tap to the active output hardware. Its
+    /// input stream format is therefore the format delivered to the I/O proc,
+    /// which can differ from the tap object's originally advertised format.
+    func readInputStreamBasicDescription() throws -> AudioStreamBasicDescription {
+        var streamsAddress = AudioObjectPropertyAddress(
+            mSelector: kAudioDevicePropertyStreams,
+            mScope: kAudioObjectPropertyScopeGlobal,
+            mElement: kAudioObjectPropertyElementMain
+        )
+        var dataSize: UInt32 = 0
+        var status = AudioObjectGetPropertyDataSize(self, &streamsAddress, 0, nil, &dataSize)
+        guard status == noErr else {
+            throw "Error reading device stream list size: \(status)"
+        }
+
+        var streamIDs = [AudioObjectID](
+            repeating: .unknown,
+            count: Int(dataSize) / MemoryLayout<AudioObjectID>.size
+        )
+        status = AudioObjectGetPropertyData(self, &streamsAddress, 0, nil, &dataSize, &streamIDs)
+        guard status == noErr else {
+            throw "Error reading device stream list: \(status)"
+        }
+
+        for streamID in streamIDs {
+            let direction: UInt32 = try streamID.read(
+                kAudioStreamPropertyDirection,
+                defaultValue: 0
+            )
+            guard direction == 1 else { continue }
+            return try streamID.read(
+                kAudioStreamPropertyVirtualFormat,
+                defaultValue: AudioStreamBasicDescription()
+            )
+        }
+
+        throw "Device has no input stream."
+    }
+
     private func requireSystemObject() throws {
         if self != .system { throw "Only supported for the system object." }
     }
@@ -307,4 +348,4 @@ extension AudioObjectID {
     func getDeviceName() throws -> String {
         return try readString(kAudioDevicePropertyDeviceNameCFString)
     }
-} 
+}

meetingnotes/ProcessTap/ProcessTap.swift

@@ -138,11 +138,11 @@ final class ProcessTap {
             tapDescription = CATapDescription(stereoMixdownOfProcesses: [process.objectID])
             logger.debug("Configuring tap for single process objectID: \(process.objectID)")
         case .systemAudio:
-            // Keep the HAL tap's buffer layout consistent with the default
+            // The transcription file is mono, so ask Core Audio for a mono
-            // output stream. AudioManager performs the stereo-to-mono mix when
+            // mixdown at the source. This avoids interpreting a stereo HAL
-            // it converts the captured audio to the 16 kHz transcription file.
+            // buffer as half as many frames before the 16 kHz conversion.
-            tapDescription = CATapDescription(stereoGlobalTapButExcludeProcesses: [])
+            tapDescription = CATapDescription(monoGlobalTapButExcludeProcesses: [])
-            logger.debug("Configuring a stereo global system audio tap.")
+            logger.info("Configuring a mono global system audio tap.")
         }
         
         tapDescription.uuid = UUID()
@@ -248,8 +248,20 @@ final class ProcessTap {
 
         do {
             logger.debug("Attempting to read audio tap stream basic description for tapID #\(tapID)...")
-            self.tapStreamDescription = try tapID.readAudioTapStreamBasicDescription()
+            let advertisedDescription = try tapID.readAudioTapStreamBasicDescription()
-            logger.debug("Successfully read tap stream description: \(String(describing: self.tapStreamDescription))")
+
+            // The aggregate device may adapt the tap to the active hardware's
+            // sample rate. Its input stream is the format actually delivered
+            // to the I/O proc, so use that rather than the tap's pre-aggregate
+            // advertisement. Using the latter can halve the written duration
+            // when, for example, a 48 kHz tap is delivered at 24 kHz.
+            do {
+                self.tapStreamDescription = try aggregateDeviceID.readInputStreamBasicDescription()
+                logger.info("Using aggregate input stream description: \(String(describing: self.tapStreamDescription), privacy: .public); tap advertised: \(String(describing: advertisedDescription), privacy: .public)")
+            } catch {
+                self.tapStreamDescription = advertisedDescription
+                logger.warning("Could not read aggregate input stream format; using tap format: \(error, privacy: .public)")
+            }
         } catch {
             logger.error("Failed to read audio tap stream basic description for tapID #\(tapID): \(error)")
             throw error // Propagate error

meetingnotes/Services/LocalAPIServer.swift

@@ -200,7 +200,9 @@ private final class LocalRecordingController {
     static let shared = LocalRecordingController()
 
     private let recordingManager = RecordingSessionManager.shared
+    private let stopGraceNanoseconds: UInt64 = 10_000_000_000
     private var isStopping = false
+    private var pendingStopTask: Task<Void, Never>?
 
     private init() {}
 
@@ -237,6 +239,12 @@ private final class LocalRecordingController {
     }
 
     func startRecording() throws -> [String: Any] {
+        if let pendingStopTask {
+            pendingStopTask.cancel()
+            self.pendingStopTask = nil
+            isStopping = false
+            return statusPayload()
+        }
         if isStopping || recordingManager.isProcessing {
             throw LocalRecordingError.processing
         }
@@ -258,14 +266,22 @@ private final class LocalRecordingController {
             return statusPayload()
         }
         isStopping = true
-        Task { [weak self] in
+        pendingStopTask = Task { [weak self] in
+            try? await Task.sleep(nanoseconds: self?.stopGraceNanoseconds ?? 0)
+            guard !Task.isCancelled else { return }
             await self?.finishRecording()
         }
         return statusPayload()
     }
 
     func cancelRecording() -> [String: Any] {
-        guard !isStopping else { return statusPayload() }
+        if let pendingStopTask {
+            pendingStopTask.cancel()
+            self.pendingStopTask = nil
+            isStopping = false
+        } else if isStopping {
+            return statusPayload()
+        }
         let meetingID = recordingManager.activeMeetingId
         recordingManager.cancelRecording()
         if let meetingID,
@@ -277,6 +293,7 @@ private final class LocalRecordingController {
     }
 
     private func finishRecording() async {
+        pendingStopTask = nil
         let meetingID = recordingManager.activeMeetingId
         let chunks = await recordingManager.stopRecording()
         guard let meetingID,