Fix inverted speaker-separation slider in diarization settings

Sources/HeardCore/CoreModels.swift

@@ -345,13 +345,15 @@ public struct AppSettings: Codable, Equatable {
     public var enableZoomDetection: Bool
     /// Detect Cisco Webex meetings.
     public var enableWebexDetection: Bool
-    /// Cosine distance threshold for speaker clustering during diarization.
-    /// Lower = stricter separation (more clusters, fewer merges); higher = looser
-    /// (fewer clusters, more chance of two voices collapsing into one). FluidAudio's
-    /// default is 0.6; we bias slightly stricter so the user can always merge
-    /// over-split speakers in the Speakers tab, which is easier than recovering
-    /// from a merged-embedding poisoning a profile.
-    public var diarizationClusteringThreshold: Double
+    /// Cosine similarity threshold for speaker clustering during diarization.
+    /// FluidAudio's AHC merges two embeddings when their cosine similarity is at or
+    /// above this value, so higher = stricter separation (more clusters, fewer
+    /// merges); lower = looser (fewer clusters, more chance of two voices
+    /// collapsing into one). FluidAudio's default is 0.6; we bias slightly
+    /// stricter so the user can always merge over-split speakers in the Speakers
+    /// tab, which is easier than recovering from a merged-embedding poisoning a
+    /// profile.
+    public var diarizationClusteringSimilarity: Double
     public var appearance: AppAppearance
     /// When enabled, preprocessing runs the app and mic tracks sequentially rather than
     /// concurrently, halving peak RAM during the VAD stage (~400 MB instead of ~800 MB).
@@ -381,7 +383,7 @@ public struct AppSettings: Codable, Equatable {
         enableTeamsDetection: true,
         enableZoomDetection: true,
         enableWebexDetection: true,
-        diarizationClusteringThreshold: 0.5,
+        diarizationClusteringSimilarity: 0.65,
         appearance: .system,
         lowMemoryMode: false
     )
@@ -406,7 +408,7 @@ public struct AppSettings: Codable, Equatable {
         enableTeamsDetection: Bool = true,
         enableZoomDetection: Bool = true,
         enableWebexDetection: Bool = true,
-        diarizationClusteringThreshold: Double = 0.5,
+        diarizationClusteringSimilarity: Double = 0.65,
         appearance: AppAppearance = .system,
         lowMemoryMode: Bool = false
     ) {
@@ -429,7 +431,7 @@ public struct AppSettings: Codable, Equatable {
         self.enableTeamsDetection = enableTeamsDetection
         self.enableZoomDetection = enableZoomDetection
         self.enableWebexDetection = enableWebexDetection
-        self.diarizationClusteringThreshold = diarizationClusteringThreshold
+        self.diarizationClusteringSimilarity = diarizationClusteringSimilarity
         self.appearance = appearance
         self.lowMemoryMode = lowMemoryMode
     }

Sources/HeardCore/Services.swift

@@ -2269,12 +2269,13 @@ public final class PipelineProcessor: ObservableObject {
             return
         }
 
-        // Configure clustering threshold so the user can bias toward more
-        // separations (lower) vs fewer (higher). The Speakers tab supports
-        // merging, but recovering from a merged embedding is harder, so we
-        // default to stricter than FluidAudio's 0.6.
-        let threshold = settingsStore.settings.diarizationClusteringThreshold
-        let config = OfflineDiarizerConfig(clusteringThreshold: threshold)
+        // FluidAudio's clusteringThreshold is a cosine *similarity* (despite the
+        // name) — AHC merges when cos_sim ≥ threshold. Higher = stricter
+        // separation (more clusters); lower = more merging. The Speakers tab
+        // supports merging, but recovering from a merged embedding is harder,
+        // so we default to stricter than FluidAudio's 0.6.
+        let similarity = settingsStore.settings.diarizationClusteringSimilarity
+        let config = OfflineDiarizerConfig(clusteringThreshold: similarity)
         let diarizer = OfflineDiarizerManager(config: config)
         try await diarizer.prepareModels()
         appDiarization = try await diarizer.process(audio: track.samples)

Sources/HeardCore/Stores.swift

@@ -160,9 +160,9 @@ public final class SettingsStore: ObservableObject {
             transcriptionModel = decoded
         }
 
-        var diarizationClusteringThreshold = base.diarizationClusteringThreshold
-        if let val = defaults.object(forKey: "diarizationClusteringThreshold") as? NSNumber {
-            diarizationClusteringThreshold = val.doubleValue
+        var diarizationClusteringSimilarity = base.diarizationClusteringSimilarity
+        if let val = defaults.object(forKey: "diarizationClusteringSimilarity") as? NSNumber {
+            diarizationClusteringSimilarity = val.doubleValue
         }
 
         settings = AppSettings(
@@ -184,7 +184,7 @@ public final class SettingsStore: ObservableObject {
             enableTeamsDetection: defaults.object(forKey: "enableTeamsDetection") as? Bool ?? base.enableTeamsDetection,
             enableZoomDetection: defaults.object(forKey: "enableZoomDetection") as? Bool ?? base.enableZoomDetection,
             enableWebexDetection: defaults.object(forKey: "enableWebexDetection") as? Bool ?? base.enableWebexDetection,
-            diarizationClusteringThreshold: diarizationClusteringThreshold
+            diarizationClusteringSimilarity: diarizationClusteringSimilarity
         )
     }
 
@@ -213,7 +213,7 @@ public final class SettingsStore: ObservableObject {
         defaults.set(settings.enableTeamsDetection, forKey: "enableTeamsDetection")
         defaults.set(settings.enableZoomDetection, forKey: "enableZoomDetection")
         defaults.set(settings.enableWebexDetection, forKey: "enableWebexDetection")
-        defaults.set(settings.diarizationClusteringThreshold, forKey: "diarizationClusteringThreshold")
+        defaults.set(settings.diarizationClusteringSimilarity, forKey: "diarizationClusteringSimilarity")
     }
 }
 

Sources/HeardCore/Views.swift

@@ -1484,36 +1484,36 @@ public struct SettingsView: View {
                                     .font(.system(size: 12, weight: .medium))
                                     .foregroundStyle(HeardTheme.Paper.ink)
                                 Spacer()
-                                Text(String(format: "%.2f", model.settingsStore.settings.diarizationClusteringThreshold))
+                                Text(String(format: "%.2f", model.settingsStore.settings.diarizationClusteringSimilarity))
                                     .font(.system(size: 12, design: .monospaced))
                                     .foregroundStyle(HeardTheme.Paper.mute)
                             }
                             HStack(spacing: 8) {
-                                Text("More speakers")
+                                Text("Fewer speakers")
                                     .font(.system(size: 10))
                                     .foregroundStyle(HeardTheme.Paper.mute)
                                 Slider(
-                                    value: settingsBinding(\.diarizationClusteringThreshold),
-                                    in: 0.30...0.80,
+                                    value: settingsBinding(\.diarizationClusteringSimilarity),
+                                    in: 0.40...0.85,
                                     step: 0.05
                                 )
-                                Text("Fewer speakers")
+                                Text("More speakers")
                                     .font(.system(size: 10))
                                     .foregroundStyle(HeardTheme.Paper.mute)
                             }
                         }
                     }
                     CardRow(isLast: false) {
-                        Text("Cosine-distance threshold for clustering voice embeddings. Lower values err on the side of splitting one person across two profiles (which you can merge in the Speakers tab); higher values may collapse two voices into one (harder to recover from). Default: 0.50.")
+                        Text("Cosine-similarity threshold for clustering voice embeddings. Higher values err on the side of splitting one person across two profiles (which you can merge in the Speakers tab); lower values may collapse two voices into one (harder to recover from). Default: 0.65.")
                             .font(.system(size: 11))
                             .foregroundStyle(HeardTheme.Paper.mute)
                     }
                     CardRow(isLast: true) {
                         HStack {
                             Spacer()
                             Button("Reset to Default") {
-                                model.settingsStore.settings.diarizationClusteringThreshold =
-                                    AppSettings.default.diarizationClusteringThreshold
+                                model.settingsStore.settings.diarizationClusteringSimilarity =
+                                    AppSettings.default.diarizationClusteringSimilarity
                             }
                             .buttonStyle(.plain)
                             .font(.system(size: 11, weight: .medium))

handoff.md

@@ -37,7 +37,7 @@ The app builds cleanly with `swift build` and runs as a menu bar app on macOS 15
 - Sequential job queue with stages: queued → preprocessing → transcribing → diarizing → assigning → complete
 - **Preprocessing**: Resample to 16kHz mono via `AudioConverter`, Silero VAD silence trimming, `VadSegmentMap` for timestamp remapping
 - **Transcription**: Parakeet TDT V2/V3 (user-selectable) via `AsrManager` with 16k sample minimum guard. Each transcribe call uses a fresh `TdtDecoderState`, so no context bleeds between tracks or jobs. Always passes `language: .english` — required by FluidAudio 0.14.x to keep v3 from emitting Cyrillic for short Latin-script utterances; ignored by v2.
-- **Diarization**: `OfflineDiarizerManager` on app track only (mic track is a single known speaker, diarization was unused). Clustering threshold is user-configurable via `AppSettings.diarizationClusteringThreshold` (default 0.50, range 0.30–0.80 in 0.05 steps; FluidAudio's library default is 0.60). Lower = stricter separation, more clusters. The 0.50 default biases toward over-splitting because merging in the Speakers tab is easier than recovering a polluted embedding.
+- **Diarization**: `OfflineDiarizerManager` on app track only (mic track is a single known speaker, diarization was unused). Clustering parameter is user-configurable via `AppSettings.diarizationClusteringSimilarity` (default 0.65, range 0.40–0.85 in 0.05 steps; FluidAudio's library default is 0.60). FluidAudio's `OfflineDiarizerConfig(clusteringThreshold:)` is actually a cosine *similarity* — AHC merges when cos_sim ≥ threshold, so higher = stricter separation, more clusters. The 0.65 default biases toward over-splitting because merging in the Speakers tab is easier than recovering a polluted embedding.
 - **Speaker Assignment**: Cosine distance matching against `SpeakerStore`, confidence margin filtering, embedding diversity management
 - Non-retryable errors (no audio, too short) fail immediately; transient errors retry 3x per session with backoff (5s, 30s, 5min) via `PipelineProcessor.executeWithRetry` (closure-driven, testable)
 - `retryCount` is cumulative across sessions with a lifetime cap of 6 (`PipelineProcessor.lifetimeRetryLimit`). User-initiated retry (`retryFailedJob`) resets `retryCount = 0` for a fresh budget.