feat(lifelog/diarize): 凝聚式合并收尾，治声纹过度切分（同人被拆成多簇）

贪心单遍聚类会把同一人切成多簇：早期噪声段一旦另起簇，后续同人段按「最近质心」各自归簇，再不会并回（实测主说话人被拆成说话人11+12+29，质心却差 0.65+）。 - 新增 SpeakerClusterer.mergeOverSplit：重聚类后做 agglomerative 合并——反复把质心余弦 ≥ mergeThreshold 的两簇并掉，质心用两簇全部成员段重算（比存的封顶滑动均值更准，能暴露被掩盖的过切：实测说话人11↔12 成员重算余弦 0.749）。约束：不并两个人工命名簇；一人工一自动并入人工保留名字；都自动并入成员多的。合并后刷新被并段的 speakerId 与相似度。 - reclusterAll / clearAllAndRecluster 收尾调用合并；mergeThreshold 默认 0.55，设置页可调 - 真机实测：32→29 簇，主说话人统一为单簇（580 段），残留仅 4 对 0.50–0.52 小簇待人工 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

feat(lifelog/diarize): 凝聚式合并收尾，治声纹过度切分（同人被拆成多簇）
贪心单遍聚类会把同一人切成多簇：早期噪声段一旦另起簇，后续同人段按「最近质心」各自归簇，再不会并回（实测主说话人被拆成说话人11+12+29，质心却差 0.65+）。 - 新增 SpeakerClusterer.mergeOverSplit：重聚类后做 agglomerative 合并——反复把质心余弦 ≥ mergeThreshold 的两簇并掉，质心用两簇全部成员段重算（比存的封顶滑动均值更准，能暴露被掩盖的过切：实测说话人11↔12 成员重算余弦 0.749）。约束：不并两个人工命名簇；一人工一自动并入人工保留名字；都自动并入成员多的。合并后刷新被并段的 speakerId 与相似度。 - reclusterAll / clearAllAndRecluster 收尾调用合并；mergeThreshold 默认 0.55，设置页可调 - 真机实测：32→29 簇，主说话人统一为单簇（580 段），残留仅 4 对 0.50–0.52 小簇待人工 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
yibo
1 parent 727ff70e
Showing 3 changed files with 92 additions and 4 deletions
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/MainActivity.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/SettingsScreen.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/asr/SpeakerClusterer.kt
@@ -864,6 +864,7 @@ private fun SpeakersTab(
                                 repo,
                                 matchThreshold = SpeakerClusterer.thresholdFromPrefs(ctx),
                                 minRecognizeMs = SpeakerClusterer.minRecognizeMsFromPrefs(ctx),
+                                mergeThreshold = SpeakerClusterer.mergeThresholdFromPrefs(ctx),
                             )
                         }
                         reclustering = false
@@ -975,6 +976,7 @@ private fun SpeakersTab(
                             repo,
                             matchThreshold = SpeakerClusterer.thresholdFromPrefs(ctx),
                             minRecognizeMs = SpeakerClusterer.minRecognizeMsFromPrefs(ctx),
+                            mergeThreshold = SpeakerClusterer.mergeThresholdFromPrefs(ctx),
                         )
                     }
                     reclustering = false
@@ -51,6 +51,7 @@ fun SettingsScreen(activity: Context) {
     var minSpeech by remember { mutableStateOf(prefs.getInt(SummaryWorker.KEY_MIN_SPEECH_SEC, SummaryWorker.DEFAULT_MIN_SPEECH_SEC)) }
     var backendId by remember { mutableStateOf(prefs.getString(SummaryWorker.KEY_BACKEND, ChatConfig.ID_2B) ?: ChatConfig.ID_2B) }
     var threshold by remember { mutableStateOf(SpeakerClusterer.thresholdFromPrefs(activity)) }
+    var mergeThreshold by remember { mutableStateOf(SpeakerClusterer.mergeThresholdFromPrefs(activity)) }
     var minRecognizeSec by remember {
         mutableStateOf(prefs.getInt(SpeakerClusterer.KEY_MIN_RECOGNIZE_SEC, (SpeakerClusterer.MIN_RECOGNIZE_MS / 1000).toInt()))
     }
@@ -98,11 +99,15 @@ fun SettingsScreen(activity: Context) {
             listOf(0.35f, 0.40f, 0.45f, 0.50f, 0.55f, 0.60f, 0.65f).map { "%.2f".format(it) to it }) {
             threshold = it; prefs.edit().putFloat(SpeakerClusterer.KEY_THRESHOLD, it).apply()
         }
+        DropdownSetting("合并阈值（越小越爱并同人）", "%.2f".format(mergeThreshold),
+            listOf(0.45f, 0.50f, 0.55f, 0.60f, 0.65f).map { "%.2f".format(it) to it }) {
+            mergeThreshold = it; prefs.edit().putFloat(SpeakerClusterer.KEY_MERGE_THRESHOLD, it).apply()
+        }
         DropdownSetting("最短识别时长（短于此不认声纹）", "$minRecognizeSec 秒",
             listOf(1, 2, 3, 5, 8).map { "$it 秒" to it }) {
             minRecognizeSec = it; prefs.edit().putInt(SpeakerClusterer.KEY_MIN_RECOGNIZE_SEC, it).apply()
         }
-        Text("改完到「声纹」tab 点「重新聚类」对历史段生效；新录音自动用新值。",
+        Text("重聚类会先贪心分簇、再把质心很近的簇按「合并阈值」并回（治同人被拆成多个）。改完到「声纹」tab 点「重新聚类」生效。",
             style = MaterialTheme.typography.labelSmall)
         HorizontalDivider()
@@ -114,11 +114,21 @@ class SpeakerClusterer(
         const val PREFS = "lifelog_prefs"
         const val KEY_THRESHOLD = "speaker_threshold"
         const val KEY_MIN_RECOGNIZE_SEC = "speaker_min_recognize_sec"
+        const val KEY_MERGE_THRESHOLD = "speaker_merge_threshold"
         fun thresholdFromPrefs(ctx: android.content.Context): Float =
             ctx.getSharedPreferences(PREFS, android.content.Context.MODE_PRIVATE)
                 .getFloat(KEY_THRESHOLD, MATCH_THRESHOLD)
+        /** 合并阈值（设置页可调）：重聚类后，质心余弦 ≥ 此值的两簇判为同一人并合并。 */
+        fun mergeThresholdFromPrefs(ctx: android.content.Context): Float =
+            ctx.getSharedPreferences(PREFS, android.content.Context.MODE_PRIVATE)
+                .getFloat(KEY_MERGE_THRESHOLD, MERGE_THRESHOLD)
+
+        /** 合并阈值默认值：贪心聚类易把同一人切成多簇（早期噪声段另起簇后互相吸段），
+         *  收尾做凝聚式合并把质心很近的簇并回来。0.55 偏保守，只并高置信对；调低并得更狠。 */
+        const val MERGE_THRESHOLD = 0.55f
+
         /** 最短识别时长（设置页可调，秒）：短于此的段完全跳过声纹识别。 */
         fun minRecognizeMsFromPrefs(ctx: android.content.Context): Long =
             ctx.getSharedPreferences(PREFS, android.content.Context.MODE_PRIVATE)
@@ -170,6 +180,7 @@ class SpeakerClusterer(
             matchThreshold: Float = MATCH_THRESHOLD,
             minNewClusterMs: Long = MIN_NEW_CLUSTER_MS,
             minRecognizeMs: Long = MIN_RECOGNIZE_MS,
+            mergeThreshold: Float = MERGE_THRESHOLD,
             maxSamplesPerCentroid: Int = MAX_SAMPLES_PER_CENTROID,
         ): Int = withContext(Dispatchers.IO) {
             // 1. 删掉自动簇，保留人工命名簇作种子。
@@ -234,8 +245,77 @@ class SpeakerClusterer(
                 repo.segments.updateSpeaker(seg.id, assignment.first, assignment.second)
             }
-            Log.i(TAG, "reclusterAll: ${segs.size} segments → ${clusters.size} clusters")
-            clusters.size
+            Log.i(TAG, "reclusterAll greedy: ${segs.size} segments → ${clusters.size} clusters")
+            // 3. 凝聚式合并：把过度切分的同人簇并回来。
+            val merges = mergeOverSplit(repo, mergeThreshold)
+            val finalCount = repo.speakers.all().size
+            Log.i(TAG, "reclusterAll: merged $merges pairs → $finalCount clusters")
+            finalCount
+        }
+
+        /**
+         * 凝聚式合并（agglomerative）：解决贪心聚类的过度切分——同一人因早期噪声段另起簇后，
+         * 后续段按「最近质心」各自归簇，再不会并回。这里反复找质心余弦最大的一对簇，
+         * ≥ [mergeThreshold] 就合并（质心用两簇**全部成员段**重算），直到没有可并的对。
+         *
+         * 约束：不合并两个**人工命名**簇（会毁掉用户标定）；一人工一自动则并入人工（保留名字）；
+         * 都自动则并入成员多的那个。合并后刷新被并段的 speakerId 与相似度。
+         * @return 合并发生的次数
+         */
+        suspend fun mergeOverSplit(
+            repo: Repository,
+            mergeThreshold: Float = MERGE_THRESHOLD,
+        ): Int = withContext(Dispatchers.IO) {
+            class WC(val id: Long, val manual: Boolean, val members: MutableList<FloatArray>, var centroid: FloatArray)
+            fun meanNorm(ms: List<FloatArray>): FloatArray {
+                val dim = ms[0].size
+                val m = FloatArray(dim)
+                for (e in ms) for (i in 0 until dim) m[i] += e[i]
+                for (i in 0 until dim) m[i] /= ms.size
+                l2normalize(m); return m
+            }
+            val wcs = ArrayList<WC>()
+            for (p in repo.speakers.all()) {
+                val mem = repo.segments.embeddingsForSpeaker(p.id).map { it.toFloatArrayLe() }
+                if (mem.isEmpty()) continue
+                wcs += WC(p.id, autoIndexOf(p.name) == null, mem.toMutableList(), meanNorm(mem))
+            }
+            var merges = 0
+            while (true) {
+                var bi = -1; var bj = -1; var best = mergeThreshold
+                for (i in wcs.indices) for (j in i + 1 until wcs.size) {
+                    if (wcs[i].manual && wcs[j].manual) continue   // 两个人工名不并
+                    val c = cosineSim(wcs[i].centroid, wcs[j].centroid)
+                    if (c >= best) { best = c; bi = i; bj = j }
+                }
+                if (bi < 0) break
+                val a = wcs[bi]; val b = wcs[bj]
+                val keep: WC; val drop: WC
+                when {
+                    a.manual && !b.manual -> { keep = a; drop = b }
+                    b.manual && !a.manual -> { keep = b; drop = a }
+                    a.members.size >= b.members.size -> { keep = a; drop = b }
+                    else -> { keep = b; drop = a }
+                }
+                keep.members.addAll(drop.members)
+                keep.centroid = meanNorm(keep.members)
+                repo.segments.reassignSpeaker(drop.id, keep.id)
+                repo.speakers.delete(drop.id)
+                wcs.remove(drop)
+                merges++
+            }
+            // 持久化保留簇的质心 + 刷新各段相似度（合并后质心变了）。
+            for (w in wcs) {
+                repo.speakers.updateEmbedding(w.id, w.centroid.toLeBytes(), minOf(w.members.size, MAX_SAMPLES_PER_CENTROID))
+            }
+            val byId = wcs.associateBy { it.id }
+            for (seg in repo.segments.allWithEmbedding()) {
+                val w = byId[seg.speakerId] ?: continue
+                val sim = cosineSim(seg.embedding!!.toFloatArrayLe(), w.centroid)
+                repo.segments.updateSpeaker(seg.id, w.id, sim)
+            }
+            Log.i(TAG, "mergeOverSplit: $merges merges, ${wcs.size} clusters remain (threshold=$mergeThreshold)")
+            merges
         }
         /**
@@ -248,11 +328,12 @@ class SpeakerClusterer(
             matchThreshold: Float = MATCH_THRESHOLD,
             minNewClusterMs: Long = MIN_NEW_CLUSTER_MS,
             minRecognizeMs: Long = MIN_RECOGNIZE_MS,
+            mergeThreshold: Float = MERGE_THRESHOLD,
         ): Int = withContext(Dispatchers.IO) {
             repo.segments.clearAllSpeakerAssignments()
             for (p in repo.speakers.all()) repo.speakers.delete(p.id)
             Log.i(TAG, "clearAllAndRecluster: wiped all profiles + assignments")
-            reclusterAll(repo, matchThreshold, minNewClusterMs, minRecognizeMs)
+            reclusterAll(repo, matchThreshold, minNewClusterMs, minRecognizeMs, mergeThreshold)
         }
     }
 }