feat: add Parakeet speaker diarization

Meetingnotes.xcodeproj/project.pbxproj

@@ -276,7 +276,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 28;
+				CURRENT_PROJECT_VERSION = 31;
 				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 = 1.1.16;
+				MARKETING_VERSION = 1.1.19;
 				ONLY_ACTIVE_ARCH = NO;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;
@@ -312,7 +312,7 @@
 				CODE_SIGN_IDENTITY = "Apple Development";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
-				CURRENT_PROJECT_VERSION = 28;
+				CURRENT_PROJECT_VERSION = 31;
 				DEVELOPMENT_ASSET_PATHS = "\"meetingnotes/Preview Content\"";
 				DEVELOPMENT_TEAM = G9LVHZAJNX;
 				ENABLE_HARDENED_RUNTIME = YES;
@@ -326,7 +326,7 @@
 					"@executable_path/../Frameworks",
 				);
 				MACOSX_DEPLOYMENT_TARGET = 15.0;
-				MARKETING_VERSION = 1.1.16;
+				MARKETING_VERSION = 1.1.19;
 				ONLY_ACTIVE_ARCH = YES;
 				OTHER_SWIFT_FLAGS = "$(inherited) -D ENABLE_TCC_SPI";
 				PRODUCT_BUNDLE_IDENTIFIER = net.jamesbone.meetingnotes;

meetingnotes/Managers/AudioManager.swift

@@ -87,8 +87,8 @@ final class AudioManager: NSObject, ObservableObject {
         }
 
         let model = UserDefaultsManager.shared.transcriptionModel
-        async let micResult = transcribe(files[0], model: model)
-        async let systemResult = transcribe(files[1], model: model)
+        async let micResult = transcribe(files[0], model: model, diarization: false)
+        async let systemResult = transcribe(files[1], model: model, diarization: true)
         let (micTranscription, systemTranscription) = await (micResult, systemResult)
         let results = [micTranscription, systemTranscription]
 
@@ -125,8 +125,8 @@ final class AudioManager: NSObject, ObservableObject {
         let model = UserDefaultsManager.shared.transcriptionModel
         let micURL = recoveryFiles.first(where: { $0.source == .mic })?.url
         let systemURL = recoveryFiles.first(where: { $0.source == .system })?.url
-        async let micResult = transcribe(micURL, model: model)
-        async let systemResult = transcribe(systemURL, model: model)
+        async let micResult = transcribe(micURL, model: model, diarization: false)
+        async let systemResult = transcribe(systemURL, model: model, diarization: true)
         let (micTranscription, systemTranscription) = await (micResult, systemResult)
         let results = [micTranscription, systemTranscription]
         let (chunks, failures) = buildTranscriptChunks(
@@ -149,10 +149,19 @@ final class AudioManager: NSObject, ObservableObject {
         lastRecoveryAudioFolderName = nil
     }
 
-    private func transcribe(_ fileURL: URL?, model: String) async -> Result<CoderAPIClient.Transcription, Error>? {
+    private func transcribe(
+        _ fileURL: URL?,
+        model: String,
+        diarization: Bool
+    ) async -> Result<CoderAPIClient.Transcription, Error>? {
         guard let fileURL else { return nil }
         do {
-            return .success(try await CoderAPIClient.shared.transcribe(fileURL: fileURL, model: model))
+            return .success(try await CoderAPIClient.shared.transcribe(
+                fileURL: fileURL,
+                model: model,
+                diarization: diarization,
+                maxSpeakerCount: 4
+            ))
         } catch {
             return .failure(error)
         }
@@ -181,6 +190,7 @@ final class AudioManager: NSObject, ObservableObject {
                         updated.append(TranscriptChunk(
                             timestamp: captureStartedAt.addingTimeInterval(max(0, segment.start)),
                             source: source,
+                            speaker: source == .system ? segment.speaker : nil,
                             text: text,
                             isFinal: true
                         ))

meetingnotes/Managers/UserDefaultsManager.swift

@@ -79,7 +79,13 @@ class UserDefaultsManager {
     }
 
     var transcriptionModel: String {
-        get { userDefaults.string(forKey: Keys.transcriptionModel) ?? "groq/whisper-large-v3-turbo" }
+        get {
+            let stored = userDefaults.string(forKey: Keys.transcriptionModel)
+            if stored == "local-whisper/whisper-large-v3-turbo" {
+                return "local-parakeet/parakeet-tdt-0.6b-v3"
+            }
+            return stored ?? "local-parakeet/parakeet-tdt-0.6b-v3"
+        }
         set { userDefaults.set(newValue, forKey: Keys.transcriptionModel) }
     }
 

meetingnotes/Models/Meeting.swift

@@ -48,30 +48,46 @@ struct TranscriptChunk: Codable, Identifiable, Hashable {
     let id: UUID
     let timestamp: Date
     let source: AudioSource
+    let speaker: Int?
     let text: String
     let isFinal: Bool
     
-    init(id: UUID = UUID(), timestamp: Date = Date(), source: AudioSource, text: String, isFinal: Bool = false) {
+    init(id: UUID = UUID(), timestamp: Date = Date(), source: AudioSource, speaker: Int? = nil, text: String, isFinal: Bool = false) {
         self.id = id
         self.timestamp = timestamp
         self.source = source
+        self.speaker = speaker
         self.text = text
         self.isFinal = isFinal
     }
+
+    var displayName: String {
+        if source == .mic { return "Me" }
+        if let speaker { return "Speaker \(speaker)" }
+        return source.displayName
+    }
 }
 
 struct CollapsedTranscriptChunk: Identifiable {
     let id: UUID
     let timestamp: Date
     let source: AudioSource
+    let speaker: Int?
     let combinedText: String
     
-    init(id: UUID = UUID(), timestamp: Date, source: AudioSource, combinedText: String) {
+    init(id: UUID = UUID(), timestamp: Date, source: AudioSource, speaker: Int? = nil, combinedText: String) {
         self.id = id
         self.timestamp = timestamp
         self.source = source
+        self.speaker = speaker
         self.combinedText = combinedText
     }
+
+    var displayName: String {
+        if source == .mic { return "Me" }
+        if let speaker { return "Speaker \(speaker)" }
+        return source.displayName
+    }
 }
 
 struct Meeting: Codable, Identifiable, Hashable {
@@ -115,7 +131,7 @@ struct Meeting: Codable, Identifiable, Hashable {
     var transcript: String {
         return transcriptChunks
             .filter { $0.isFinal }
-            .map { "[\($0.source.rawValue)] \($0.text)" }
+            .map { "[\($0.displayName)] \($0.text)" }
             .joined(separator: " ")
     }
     
@@ -124,7 +140,7 @@ struct Meeting: Codable, Identifiable, Hashable {
         let finalChunks = transcriptChunks.filter { $0.isFinal }
 
         return finalChunks.map { chunk in
-            "[\(TranscriptTimestampFormatter.string(from: chunk.timestamp))] \(chunk.source.copyPrefix): \(chunk.text)"
+            "[\(TranscriptTimestampFormatter.string(from: chunk.timestamp))] \(chunk.displayName): \(chunk.text)"
         }.joined(separator: "\n")
     }
     
@@ -135,6 +151,7 @@ struct Meeting: Codable, Identifiable, Hashable {
                 id: chunk.id,
                 timestamp: chunk.timestamp,
                 source: chunk.source,
+                speaker: chunk.speaker,
                 combinedText: chunk.text
             )
         }

meetingnotes/ProcessTap/ProcessTap.swift

@@ -138,8 +138,11 @@ final class ProcessTap {
             tapDescription = CATapDescription(stereoMixdownOfProcesses: [process.objectID])
             logger.debug("Configuring tap for single process objectID: \(process.objectID)")
         case .systemAudio:
-            tapDescription = CATapDescription(monoGlobalTapButExcludeProcesses: [])
-            logger.debug("Configuring a global system audio tap.")
+            // Keep the HAL tap's buffer layout consistent with the default
+            // output stream. AudioManager performs the stereo-to-mono mix when
+            // it converts the captured audio to the 16 kHz transcription file.
+            tapDescription = CATapDescription(stereoGlobalTapButExcludeProcesses: [])
+            logger.debug("Configuring a stereo global system audio tap.")
         }
         
         tapDescription.uuid = UUID()

meetingnotes/Services/CoderAPIClient.swift

@@ -1,3 +1,4 @@
+import AVFoundation
 import Foundation
 
 struct CoderModel: Codable, Identifiable, Hashable {
@@ -20,6 +21,7 @@ struct CoderModel: Codable, Identifiable, Hashable {
 
     var supportsChat: Bool { capabilities.isEmpty || capabilities.contains("chat") }
     var supportsTranscription: Bool { capabilities.contains("audio_transcription") }
+    var supportsSpeakerDiarization: Bool { capabilities.contains("speaker_diarization") }
 }
 
 enum CoderAPIError: LocalizedError {
@@ -53,6 +55,7 @@ final class CoderAPIClient {
             let start: TimeInterval
             let end: TimeInterval
             let text: String
+            let speaker: Int?
         }
 
         let text: String
@@ -69,10 +72,25 @@ final class CoderAPIClient {
     }
 
     private struct TranscriptionResponse: Decodable {
+        struct Word: Decodable {
+            let word: String
+            let start: TimeInterval
+            let end: TimeInterval
+            let speaker: Int?
+        }
+
         let text: String
         let segments: [Transcription.Segment]?
+        let words: [Word]?
     }
 
+    private struct AudioChunk {
+        let url: URL
+        let offset: TimeInterval
+        let isTemporary: Bool
+    }
+
+    private let transcriptionChunkDuration: TimeInterval = 3 * 60
     private let transcriptionSession: URLSession
 
     private init() {
@@ -151,15 +169,85 @@ final class CoderAPIClient {
         }
     }
 
-    func transcribe(fileURL: URL, model: String, language: String = "en") async throws -> Transcription {
+    func transcribe(
+        fileURL: URL,
+        model: String,
+        language: String = "en",
+        diarization: Bool = false,
+        maxSpeakerCount: Int = 4
+    ) async throws -> Transcription {
         let selectedModel = model.trimmingCharacters(in: .whitespacesAndNewlines)
         guard !selectedModel.isEmpty else { throw CoderAPIError.missingModel("transcription") }
         let apiKey = try requiredAPIKey(KeychainHelper.shared.getCoderAPIKey() ?? "")
+        let chunks = try makeAudioChunks(from: fileURL, preserveSpeakerIdentity: diarization)
+        defer {
+            for chunk in chunks where chunk.isTemporary {
+                try? FileManager.default.removeItem(at: chunk.url)
+            }
+        }
+
+        var textParts: [String] = []
+        var segments: [Transcription.Segment] = []
+        var lastNormalizedText = ""
+        var consecutiveDuplicateCount = 0
+
+        for chunk in chunks {
+            let transcription = try await transcribeChunk(
+                chunk.url,
+                model: selectedModel,
+                language: language,
+                apiKey: apiKey,
+                diarization: diarization,
+                maxSpeakerCount: maxSpeakerCount
+            )
+            if transcription.segments.isEmpty {
+                let text = transcription.text.trimmingCharacters(in: .whitespacesAndNewlines)
+                if !text.isEmpty {
+                    textParts.append(text)
+                    segments.append(.init(start: chunk.offset, end: chunk.offset, text: text, speaker: nil))
+                }
+                continue
+            }
+
+            for segment in transcription.segments {
+                let text = segment.text.trimmingCharacters(in: .whitespacesAndNewlines)
+                guard !text.isEmpty else { continue }
+                let normalized = text.lowercased()
+                if normalized == lastNormalizedText {
+                    consecutiveDuplicateCount += 1
+                } else {
+                    lastNormalizedText = normalized
+                    consecutiveDuplicateCount = 1
+                }
+                guard consecutiveDuplicateCount <= 2 else { continue }
+                textParts.append(text)
+                segments.append(.init(
+                    start: segment.start + chunk.offset,
+                    end: segment.end + chunk.offset,
+                    text: text,
+                    speaker: segment.speaker
+                ))
+            }
+        }
+
+        return Transcription(text: textParts.joined(separator: "\n"), segments: segments)
+    }
+
+    private func transcribeChunk(
+        _ fileURL: URL,
+        model: String,
+        language: String,
+        apiKey: String,
+        diarization: Bool,
+        maxSpeakerCount: Int
+    ) async throws -> Transcription {
         let boundary = "Meetingnotes-\(UUID().uuidString)"
         let bodyURL = try makeMultipartBody(
             audioURL: fileURL,
-            model: selectedModel,
+            model: model,
             language: language,
+            diarization: diarization,
+            maxSpeakerCount: maxSpeakerCount,
             boundary: boundary
         )
         defer { try? FileManager.default.removeItem(at: bodyURL) }
@@ -175,7 +263,125 @@ final class CoderAPIClient {
         let (data, response) = try await transcriptionSession.upload(for: request, fromFile: bodyURL)
         try validate(response: response, data: data)
         let decoded = try JSONDecoder().decode(TranscriptionResponse.self, from: data)
-        return Transcription(text: decoded.text, segments: decoded.segments ?? [])
+        let segments = decoded.segments ?? segments(from: decoded.words ?? [])
+        return Transcription(text: decoded.text, segments: segments)
+    }
+
+    private func makeAudioChunks(from fileURL: URL, preserveSpeakerIdentity: Bool) throws -> [AudioChunk] {
+        let input = try AVAudioFile(forReading: fileURL)
+        let format = input.processingFormat
+        guard format.sampleRate > 0 else { throw CoderAPIError.invalidResponse }
+
+        let framesPerChunk = preserveSpeakerIdentity
+            ? max(1, input.length)
+            : AVAudioFramePosition(format.sampleRate * transcriptionChunkDuration)
+        var chunks: [AudioChunk] = []
+        var frameOffset: AVAudioFramePosition = 0
+
+        do {
+            while frameOffset < input.length {
+                let frameCount = min(framesPerChunk, input.length - frameOffset)
+                let chunkURL = FileManager.default.temporaryDirectory
+                    .appendingPathComponent("meetingnotes-transcription-\(UUID().uuidString).wav")
+                try writeAudioChunk(
+                    from: input,
+                    frameCount: frameCount,
+                    format: format,
+                    to: chunkURL
+                )
+                chunks.append(AudioChunk(
+                    url: chunkURL,
+                    offset: Double(frameOffset) / format.sampleRate,
+                    isTemporary: true
+                ))
+                frameOffset += frameCount
+            }
+            return chunks
+        } catch {
+            for chunk in chunks {
+                try? FileManager.default.removeItem(at: chunk.url)
+            }
+            throw error
+        }
+    }
+
+    private func writeAudioChunk(
+        from input: AVAudioFile,
+        frameCount: AVAudioFramePosition,
+        format: AVAudioFormat,
+        to outputURL: URL
+    ) throws {
+        let settings: [String: Any] = [
+            AVFormatIDKey: kAudioFormatLinearPCM,
+            AVSampleRateKey: format.sampleRate,
+            AVNumberOfChannelsKey: format.channelCount,
+            AVLinearPCMBitDepthKey: 16,
+            AVLinearPCMIsFloatKey: false,
+            AVLinearPCMIsBigEndianKey: false,
+            AVLinearPCMIsNonInterleaved: false
+        ]
+        let output = try AVAudioFile(
+            forWriting: outputURL,
+            settings: settings,
+            commonFormat: format.commonFormat,
+            interleaved: format.isInterleaved
+        )
+        var remaining = frameCount
+        while remaining > 0 {
+            let requestedFrames = AVAudioFrameCount(min(remaining, 8_192))
+            guard let buffer = AVAudioPCMBuffer(pcmFormat: format, frameCapacity: requestedFrames) else {
+                throw CoderAPIError.invalidResponse
+            }
+            try input.read(into: buffer, frameCount: requestedFrames)
+            guard buffer.frameLength > 0 else { break }
+            try output.write(from: buffer)
+            remaining -= AVAudioFramePosition(buffer.frameLength)
+        }
+    }
+
+    private func segments(from words: [TranscriptionResponse.Word]) -> [Transcription.Segment] {
+        var result: [Transcription.Segment] = []
+        var currentWords: [String] = []
+        var currentStart: TimeInterval?
+        var currentEnd: TimeInterval = 0
+        var currentSpeaker: Int?
+
+        func flush() {
+            guard let start = currentStart, !currentWords.isEmpty else { return }
+            result.append(.init(
+                start: start,
+                end: currentEnd,
+                text: currentWords.joined(separator: " "),
+                speaker: currentSpeaker
+            ))
+            currentWords.removeAll(keepingCapacity: true)
+            currentStart = nil
+            currentEnd = 0
+            currentSpeaker = nil
+        }
+
+        for word in words {
+            let text = word.word.trimmingCharacters(in: .whitespacesAndNewlines)
+            guard !text.isEmpty else { continue }
+            let speakerChanged = currentStart != nil && word.speaker != currentSpeaker
+            let longPause = currentStart != nil && word.start - currentEnd > 1.5
+            if speakerChanged || longPause { flush() }
+
+            if currentStart == nil {
+                currentStart = word.start
+                currentSpeaker = word.speaker
+            }
+            currentWords.append(text)
+            currentEnd = word.end
+
+            let sentenceEnded = text.last.map { ".!?".contains($0) } ?? false
+            let duration = currentEnd - (currentStart ?? currentEnd)
+            if currentWords.count >= 40 || (sentenceEnded && (currentWords.count >= 12 || duration >= 8)) {
+                flush()
+            }
+        }
+        flush()
+        return result
     }
 
     private func endpoint(baseURL: String, path: String) throws -> URL {
@@ -207,7 +413,14 @@ final class CoderAPIClient {
         }
     }
 
-    private func makeMultipartBody(audioURL: URL, model: String, language: String, boundary: String) throws -> URL {
+    private func makeMultipartBody(
+        audioURL: URL,
+        model: String,
+        language: String,
+        diarization: Bool,
+        maxSpeakerCount: Int,
+        boundary: String
+    ) throws -> URL {
         let bodyURL = FileManager.default.temporaryDirectory.appendingPathComponent("meetingnotes-upload-\(UUID().uuidString).body")
         _ = FileManager.default.createFile(atPath: bodyURL.path, contents: nil)
         let output = try FileHandle(forWritingTo: bodyURL)
@@ -219,7 +432,11 @@ final class CoderAPIClient {
         try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"model\"\r\n\r\n\(model)\r\n")
         try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"language\"\r\n\r\n\(language)\r\n")
         try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"response_format\"\r\n\r\nverbose_json\r\n")
-        try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"file\"; filename=\"\(audioURL.lastPathComponent)\"\r\nContent-Type: audio/mp4\r\n\r\n")
+        if diarization {
+            try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"diarization\"\r\n\r\ntrue\r\n")
+            try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"max_speaker_count\"\r\n\r\n\(maxSpeakerCount)\r\n")
+        }
+        try write("--\(boundary)\r\nContent-Disposition: form-data; name=\"file\"; filename=\"\(audioURL.lastPathComponent)\"\r\nContent-Type: audio/wav\r\n\r\n")
         let input = try FileHandle(forReadingFrom: audioURL)
         defer { try? input.close() }
         while let chunk = try input.read(upToCount: 1 << 20), !chunk.isEmpty {

meetingnotes/Views/MeetingListView.swift

@@ -208,12 +208,12 @@ struct CollapsedTranscriptChunkView: View {
                     .font(.caption)
                     .foregroundColor(chunk.source == .mic ? .blue : .orange)
                 
-                Text(chunk.source.displayName)
+                Text(chunk.displayName)
                     .font(.caption)
                     .fontWeight(.medium)
                     .foregroundColor(chunk.source == .mic ? .blue : .orange)
             }
-            .frame(width: 50, alignment: .leading)
+            .frame(width: 78, alignment: .leading)
             
             // Transcript text
             Text(chunk.combinedText)

meetingnotes/Views/SettingsView.swift

@@ -58,6 +58,12 @@ struct SettingsView: View {
                                 Text(model.displayName).tag(model.id)
                             }
                         }
+
+                        if viewModel.coderModels.first(where: { $0.id == viewModel.settings.transcriptionModel })?.supportsSpeakerDiarization == true {
+                            Label("Remote participants are labeled Speaker 1–4; your microphone is labeled Me.", systemImage: "person.2.wave.2")
+                                .font(.caption)
+                                .foregroundColor(.secondary)
+                        }
                     }
 
                     Text("The token is stored locally in Keychain. Audio and note generation are sent only to this Coder service.")