feat(lifelog-recorder): v0.3 — 声纹识别 + FTS4 中文搜索

声纹（用户语音里提了 3 次的 feature）： - SpeakerEmbedder.kt: sherpa-onnx SpeakerEmbeddingExtractor 包装 3dspeaker CAMPPlus 16k zh-cn 模型（28 MB int8 ONNX），192 维 embedding - ModelDownloader.ensureSpeakerModel(): 首次启动从 hf-mirror.com 自动下 - Segment 实体加 embedding/speakerId/speakerSim 三列 - SpeakerProfile 实体（id/name/embedding/sampleCount/createdAt）+ SpeakerDao - RecorderService.handleSegment 每段算 embedding + 与已注册 profiles 算 cosine，threshold ≥0.5 才 assign - RecorderService 通过 speakers.observeAll() 实时 reload profile 缓存 - MainActivity 点时间轴段 → AlertDialog 输入名字 → 创建 SpeakerProfile - TimelineRow 显示 "name (sim=0.xx)" 或 "—" / "未匹配 (0.xx)" FTS4 全文搜索： - TranscriptFts 实体用 @Fts4(contentEntity=Transcript, tokenizer="unicode61") Room 自动建虚表 + 同步触发器，unicode61 把中文字当 token - searchTimeline DAO 查询 JOIN transcript_fts MATCH :query - MainActivity 顶部 OutlinedTextField 实时切换 timelineFlow - sanitizeFtsQuery 给每 token 加 * 做前缀匹配，剥掉 FTS metachars Schema: - AppDatabase v1 → v2 + fallbackToDestructiveMigration（v0.3 dev 数据可丢） - 新增 SpeakerProfile + TranscriptFts 虚表 + Segment 三列 HANDOFF.md §五更新：声纹 + 搜索移到「已完成」，下一档变成 LLM 纪要 worker / 存储 TTL / 段播放 / 多 anchor 平均 / lemonPCRM 回链。实测（Pixel 8 Pro Tensor G3，Android 16）： - speaker model load: 441 ms - embedding 计算开销 ~10-30 ms/段，跟 paraformer 同期完成 - enroll "Me" anchor → 后续段在 logcat 见 "speaker=Me sim=0.xx" Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

feat(lifelog-recorder): v0.3 — 声纹识别 + FTS4 中文搜索
声纹（用户语音里提了 3 次的 feature）： - SpeakerEmbedder.kt: sherpa-onnx SpeakerEmbeddingExtractor 包装 3dspeaker CAMPPlus 16k zh-cn 模型（28 MB int8 ONNX），192 维 embedding - ModelDownloader.ensureSpeakerModel(): 首次启动从 hf-mirror.com 自动下 - Segment 实体加 embedding/speakerId/speakerSim 三列 - SpeakerProfile 实体（id/name/embedding/sampleCount/createdAt）+ SpeakerDao - RecorderService.handleSegment 每段算 embedding + 与已注册 profiles 算 cosine，threshold ≥0.5 才 assign - RecorderService 通过 speakers.observeAll() 实时 reload profile 缓存 - MainActivity 点时间轴段 → AlertDialog 输入名字 → 创建 SpeakerProfile - TimelineRow 显示 "name (sim=0.xx)" 或 "—" / "未匹配 (0.xx)" FTS4 全文搜索： - TranscriptFts 实体用 @Fts4(contentEntity=Transcript, tokenizer="unicode61") Room 自动建虚表 + 同步触发器，unicode61 把中文字当 token - searchTimeline DAO 查询 JOIN transcript_fts MATCH :query - MainActivity 顶部 OutlinedTextField 实时切换 timelineFlow - sanitizeFtsQuery 给每 token 加 * 做前缀匹配，剥掉 FTS metachars Schema: - AppDatabase v1 → v2 + fallbackToDestructiveMigration（v0.3 dev 数据可丢） - 新增 SpeakerProfile + TranscriptFts 虚表 + Segment 三列 HANDOFF.md §五更新：声纹 + 搜索移到「已完成」，下一档变成 LLM 纪要 worker / 存储 TTL / 段播放 / 多 anchor 平均 / lemonPCRM 回链。实测（Pixel 8 Pro Tensor G3，Android 16）： - speaker model load: 441 ms - embedding 计算开销 ~10-30 ms/段，跟 paraformer 同期完成 - enroll "Me" anchor → 后续段在 logcat 见 "speaker=Me sim=0.xx" Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
yibo
1 parent d67ce3d2
Showing 9 changed files with 512 additions and 38 deletions
lifelog-recorder/HANDOFF.md
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/MainActivity.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/asr/ModelDownloader.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/asr/SpeakerEmbedder.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/data/AppDatabase.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/data/Daos.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/data/Entities.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/data/Repository.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/service/RecorderService.kt
@@ -40,7 +40,7 @@ APK 大小:    62 MB（含 34 MB sherpa-onnx .so + 0.6 MB silero VAD + 3 MB samp
 ```
 app/src/main/kotlin/com/lemon/lifelog/
 ├── LifelogApp.kt                — Application，建 NotificationChannel
-├── MainActivity.kt              — Compose UI，时间轴 + 电池豁免 UI + 测试按钮
+├── MainActivity.kt              — Compose UI，时间轴 + 搜索 + 声纹 enroll dialog
 ├── service/
 │   ├── RecorderService.kt       — 前台服务，pipeline 主体 ←★ 改 ASR 流程改这里
 │   ├── RecorderState.kt         — StateFlow 把服务状态推到 UI
@@ -50,12 +50,13 @@ app/src/main/kotlin/com/lemon/lifelog/
 │   └── SegmentWriter.kt         — Opus encode via MediaCodec + Ogg muxer，~17× 压缩
 ├── asr/
 │   ├── ParaformerEngine.kt      — sherpa-onnx paraformer 包装 ←★ 主 ASR
-│   ├── ModelDownloader.kt       — 首次启动下模型 + 断点续传
+│   ├── SpeakerEmbedder.kt       — sherpa-onnx 3dspeaker campplus 192-d 嵌入 + cosineSim
+│   ├── ModelDownloader.kt       — 首次启动下模型（paraformer 232 MB + speaker 28 MB）
 │   └── XfyunIatClient.kt        — 讯飞大模型 WS 客户端，仅 fallback / 云复核用
 └── data/
-    ├── Entities.kt              — Segment / Transcript / Summary
-    ├── Daos.kt                  — 含 observeTimeline() 给 UI 流
-    ├── AppDatabase.kt           — Room 1.0
+    ├── Entities.kt              — Segment / Transcript / Summary / SpeakerProfile / TranscriptFts
+    ├── Daos.kt                  — observeTimeline + searchTimeline (FTS4) + SpeakerDao
+    ├── AppDatabase.kt           — Room v2，destructive migration（v0.3 schema 大改）
     └── Repository.kt
 ```
@@ -151,17 +152,18 @@ VAD 参数（在 `RecorderService.runPipeline` 里调）：
 ## §五 §六 待办（按优先级）
-✅ **已完成（v0.2，2026-06）：Opus 编码 / BootReceiver / 电池白名单**
+✅ **已完成**：
+- v0.2 (2026-06)：Opus 编码 / BootReceiver / 电池白名单
+- v0.3 (2026-06)：声纹识别（3dspeaker campplus）/ FTS4 中文搜索
 剩下：
-1. **声纹识别 / speaker diarization**（🔥 用户当面提过 3 次）：sherpa-onnx 已有 `Speaker.kt` + `SpeakerEmbeddingExtractorConfig.kt`。流程：(a)「这是我」录音 → 建 anchor embedding 存 SQLite；(b) 每段算 embedding 算 cosine 找最近邻；(c) Transcript 加 `speakerHint` 列。需要下 `sherpa-onnx-3dspeaker-zh.onnx`（~38 MB）。
-2. **FTS5 中文搜索**：Room + sqlite-jsqlite-with-icu。Schema 加 `TranscriptFts` 虚表用 ICU tokenizer。UI 加搜索框 + 高亮命中段。
-3. **LLM 纪要 worker**：WorkManager cron。每小时 / 静音 30min 触发，按时间窗 group segments → POST 到 hermes gateway `http://127.0.0.1:8642/v1/chat/completions`（需要先把 hermes api_server platform 配置起来，见 `~/.hermes/config.yaml`）。
-4. **存储 TTL**：删超过 N 天的 .opus，文字永留。Periodic WorkManager job。
-5. **设置页**：存储用量、模型切换、provider 切换、暂停时段配置（v1 定的 MVP 是「一键暂停就够」，但 v2 可能要细化）。
-6. **段播放**：时间轴每条点开能直接播原音（Opus 系统原生支持，MediaPlayer 就能放）。
-7. **回链 lemonPCRM**：Transcript 提到的人名 / 项目 → GraphStore `MENTIONED_IN` 边。
+1. **LLM 纪要 worker**：WorkManager cron。每小时 / 静音 30min 触发，按时间窗 group segments → POST 到 hermes gateway `http://127.0.0.1:8642/v1/chat/completions`（需要先把 hermes api_server platform 配置起来，见 `~/.hermes/config.yaml`）。
+2. **存储 TTL**：删超过 N 天的 .opus，文字永留。Periodic WorkManager job。
+3. **设置页**：存储用量、模型切换、provider 切换、暂停时段配置（v1 定的 MVP 是「一键暂停就够」，但 v2 可能要细化）。
+4. **段播放**：时间轴每条点开能直接播原音（Opus 系统原生支持，MediaPlayer 就能放）。当前点开是「记录说话人」对话框，需要分离 tap vs long-press 或加二级菜单。
+5. **多 anchor 平均 / 声纹回填**：现在 enroll 是单段单 anchor，可以「连选多段都标 我」自动算平均 embedding；并对老段做一次性 backfill 匹配。
+6. **回链 lemonPCRM**：Transcript 提到的人名 / 项目 → GraphStore `MENTIONED_IN` 边。
 ### 存储预算（v0.2 后）
@@ -13,6 +13,7 @@ import androidx.activity.ComponentActivity
 import androidx.activity.compose.rememberLauncherForActivityResult
 import androidx.activity.compose.setContent
 import androidx.activity.result.contract.ActivityResultContracts
+import androidx.compose.foundation.clickable
 import androidx.compose.foundation.layout.Arrangement
 import androidx.compose.foundation.layout.Column
 import androidx.compose.foundation.layout.Row
@@ -25,12 +26,17 @@ import androidx.compose.foundation.layout.padding
 import androidx.compose.foundation.lazy.LazyColumn
 import androidx.compose.foundation.lazy.items
 import androidx.compose.foundation.rememberScrollState
+import androidx.compose.foundation.text.KeyboardOptions
 import androidx.compose.foundation.verticalScroll
+import androidx.compose.material3.AlertDialog
 import androidx.compose.material3.Button
 import androidx.compose.material3.HorizontalDivider
 import androidx.compose.material3.MaterialTheme
+import androidx.compose.material3.OutlinedTextField
 import androidx.compose.material3.Surface
 import androidx.compose.material3.Text
+import androidx.compose.material3.TextButton
+import androidx.compose.material3.TextField
 import androidx.compose.runtime.Composable
 import androidx.compose.runtime.collectAsState
 import androidx.compose.runtime.DisposableEffect
@@ -47,10 +53,15 @@ import androidx.compose.ui.unit.dp
 import androidx.core.content.ContextCompat
 import com.lemon.lifelog.asr.ModelDownloader
 import com.lemon.lifelog.asr.ParaformerEngine
+import com.lemon.lifelog.asr.SpeakerEmbedder
 import com.lemon.lifelog.asr.XfyunIatClient
+import com.lemon.lifelog.asr.toFloatArrayLe
+import com.lemon.lifelog.asr.toLeBytes
 import com.lemon.lifelog.asr.transcribeWithStats
 import com.lemon.lifelog.data.Repository
+import com.lemon.lifelog.data.Segment
 import com.lemon.lifelog.data.SegmentRow
+import com.lemon.lifelog.data.SpeakerProfile
 import com.lemon.lifelog.service.RecorderService
 import com.lemon.lifelog.service.RecorderState
 import kotlinx.coroutines.flow.Flow
@@ -90,9 +101,19 @@ private fun MainScreen() {
     val activity = androidx.compose.ui.platform.LocalContext.current
     val scope = rememberCoroutineScope()
-    // Timeline: observe segments + transcripts from Room.
+    // Timeline + search.
     val repo = remember(activity) { Repository(activity) }
-    val timeline by repo.segments.observeTimeline(200).collectAsState(initial = emptyList())
+    var searchQuery by remember { mutableStateOf("") }
+    val effectiveSearch = remember(searchQuery) { sanitizeFtsQuery(searchQuery) }
+    val timelineFlow = remember(effectiveSearch) {
+        if (effectiveSearch.isEmpty()) repo.segments.observeTimeline(200)
+        else repo.segments.searchTimeline(effectiveSearch, 200)
+    }
+    val timeline by timelineFlow.collectAsState(initial = emptyList())
+    val speakers by repo.speakers.observeAll().collectAsState(initial = emptyList())
+
+    // Enrollment dialog state.
+    var enrollFor by remember { mutableStateOf<SegmentRow?>(null) }
     // Battery optimisation exemption — recompute whenever activity resumes
     // (user may have just returned from the system settings dialog).
@@ -165,14 +186,32 @@ private fun MainScreen() {
         HorizontalDivider()
-        Text("时间轴（${timeline.size}）", style = MaterialTheme.typography.titleMedium)
+        Row(verticalAlignment = androidx.compose.ui.Alignment.CenterVertically,
+            horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            Text(
+                if (searchQuery.isEmpty()) "时间轴（${timeline.size}）"
+                else "搜索结果（${timeline.size}）",
+                style = MaterialTheme.typography.titleMedium,
+            )
+            Text("· 声纹库 ${speakers.size}",
+                style = MaterialTheme.typography.bodySmall)
+        }
+        OutlinedTextField(
+            value = searchQuery,
+            onValueChange = { searchQuery = it },
+            modifier = Modifier.fillMaxWidth(),
+            placeholder = { Text("中文搜索…") },
+            singleLine = true,
+        )
         LazyColumn(
             modifier = Modifier
                 .fillMaxWidth()
                 .heightIn(min = 200.dp, max = 600.dp),
             verticalArrangement = Arrangement.spacedBy(4.dp),
         ) {
-            items(timeline, key = { it.id }) { row -> TimelineRow(row) }
+            items(timeline, key = { it.id }) { row ->
+                TimelineRow(row, onClick = { enrollFor = row })
+            }
         }
         HorizontalDivider()
@@ -207,6 +246,114 @@ private fun MainScreen() {
         Spacer(Modifier.height(8.dp))
         Text(asrLog, style = MaterialTheme.typography.bodySmall)
     }
+
+    // Speaker enrollment dialog.
+    enrollFor?.let { row ->
+        EnrollSpeakerDialog(
+            row = row,
+            existingNames = speakers.map { it.name },
+            onDismiss = { enrollFor = null },
+            onConfirm = { name ->
+                val target = enrollFor
+                enrollFor = null
+                if (target != null && name.isNotBlank()) {
+                    scope.launch {
+                        enrollSpeakerForSegment(activity, repo, target.id, name)
+                    }
+                }
+            },
+        )
+    }
+}
+
+@Composable
+private fun EnrollSpeakerDialog(
+    row: SegmentRow,
+    existingNames: List<String>,
+    onDismiss: () -> Unit,
+    onConfirm: (String) -> Unit,
+) {
+    var name by remember { mutableStateOf("") }
+    AlertDialog(
+        onDismissRequest = onDismiss,
+        title = { Text("记录说话人") },
+        text = {
+            Column(verticalArrangement = Arrangement.spacedBy(8.dp)) {
+                Text(
+                    "用这一段的声音建一个声纹 anchor。之后说话风格相近的段会自动打上同名标签。",
+                    style = MaterialTheme.typography.bodySmall,
+                )
+                Text(
+                    "段 #${row.id}: ${row.text?.take(40) ?: "(无转写)"}",
+                    style = MaterialTheme.typography.bodyMedium,
+                )
+                OutlinedTextField(
+                    value = name,
+                    onValueChange = { name = it },
+                    label = { Text("说话人名字（如：我、老婆、张三）") },
+                    singleLine = true,
+                )
+                if (existingNames.isNotEmpty()) {
+                    Text(
+                        "已存声纹：${existingNames.joinToString("、")}",
+                        style = MaterialTheme.typography.labelSmall,
+                    )
+                }
+            }
+        },
+        confirmButton = {
+            TextButton(onClick = { onConfirm(name) }, enabled = name.isNotBlank()) {
+                Text("保存")
+            }
+        },
+        dismissButton = {
+            TextButton(onClick = onDismiss) { Text("取消") }
+        },
+    )
+}
+
+/**
+ * Read embedding from segment row, save as new SpeakerProfile.
+ * Future segments will be auto-matched against it (see RecorderService.refreshProfiles).
+ */
+private suspend fun enrollSpeakerForSegment(
+    ctx: Context,
+    repo: Repository,
+    segmentId: Long,
+    name: String,
+) {
+    withContext(Dispatchers.IO) {
+        val seg = repo.segments.get(segmentId) ?: return@withContext
+        val embedding = seg.embedding ?: return@withContext
+        repo.speakers.insert(
+            SpeakerProfile(
+                name = name.trim(),
+                embedding = embedding,
+                sampleCount = 1,
+            ),
+        )
+        android.util.Log.i("MainActivity", "enrolled speaker '$name' from segment #$segmentId")
+    }
+}
+
+/**
+ * FTS4 MATCH expects a query string.  For Chinese character-level search with the
+ * unicode61 tokenizer, append `*` to each token so substring/prefix lookup works.
+ * Strip FTS metacharacters that would otherwise crash the parser.
+ */
+private fun sanitizeFtsQuery(raw: String): String {
+    val cleaned = raw.trim()
+        .replace("\"", "")
+        .replace("'", "")
+        .replace("(", "")
+        .replace(")", "")
+        .replace(":", "")
+    if (cleaned.isEmpty()) return ""
+    // Treat whole input as one prefix term.  For longer queries the user can put spaces
+    // for AND-of-terms; we add * suffix to each so partial matches work.
+    return cleaned.split(Regex("\\s+"))
+        .filter { it.isNotEmpty() }
+        .joinToString(" ") { "$it*" }
 }
 private fun isBatteryOptExempt(ctx: Context): Boolean {
@@ -226,12 +373,23 @@ private fun requestBatteryOptExemption(ctx: Context) {
 private val TIMELINE_FMT = SimpleDateFormat("HH:mm:ss", Locale.US)
 @Composable
-private fun TimelineRow(row: SegmentRow) {
+private fun TimelineRow(row: SegmentRow, onClick: () -> Unit) {
     val time = TIMELINE_FMT.format(Date(row.startTs))
     val durSec = "%.1fs".format(row.durationMs / 1000.0)
-    Column(modifier = Modifier.fillMaxWidth().padding(vertical = 2.dp)) {
+    val speakerLabel = when {
+        row.speakerName != null && row.speakerSim != null ->
+            "${row.speakerName} (${"%.2f".format(row.speakerSim)})"
+        row.speakerSim != null -> "未匹配 (${"%.2f".format(row.speakerSim)})"
+        else -> "—"
+    }
+    Column(
+        modifier = Modifier
+            .fillMaxWidth()
+            .clickable(onClick = onClick)
+            .padding(vertical = 4.dp),
+    ) {
         Text(
-            "$time  ·  $durSec  ·  #${row.id}",
+            "$time  ·  $durSec  ·  #${row.id}  ·  $speakerLabel",
             style = MaterialTheme.typography.labelSmall,
         )
         Text(
@@ -24,6 +24,8 @@ class ModelDownloader(private val ctx: Context) {
         val present: Boolean,
     )
+    data class SpeakerModelFile(val path: String, val present: Boolean)
+
     fun paths(): ModelFiles {
         val dir = File(ctx.filesDir, "models/paraformer-zh").apply { mkdirs() }
         val model = File(dir, "model.int8.onnx")
@@ -32,16 +34,21 @@ class ModelDownloader(private val ctx: Context) {
         return ModelFiles(model.absolutePath, tokens.absolutePath, present)
     }
+    fun speakerPath(): SpeakerModelFile {
+        val dir = File(ctx.filesDir, "models/3dspeaker").apply { mkdirs() }
+        val model = File(dir, "campplus_zh.onnx")
+        // expected ~28 MB, accept anything > 20 MB as valid
+        val present = model.exists() && model.length() > 20_000_000L
+        return SpeakerModelFile(model.absolutePath, present)
+    }
+
     suspend fun ensure(progress: (String) -> Unit): ModelFiles = withContext(Dispatchers.IO) {
         val state = paths()
         if (state.present) {
-            progress("model present (${File(state.modelPath).length() / 1_000_000}MB), skip")
+            progress("paraformer model present (${File(state.modelPath).length() / 1_000_000}MB), skip")
             return@withContext state
         }
-        val client = OkHttpClient.Builder()
-            .connectTimeout(15, TimeUnit.SECONDS)
-            .readTimeout(120, TimeUnit.SECONDS)
-            .build()
+        val client = http()
         val tasks = listOf(
             DownloadTask(
@@ -68,6 +75,35 @@ class ModelDownloader(private val ctx: Context) {
         paths()
     }
+    /**
+     * 3dspeaker CAMPPlus 16k speaker embedding (192 dim) from sherpa-onnx releases.
+     * ~28 MB on-device.
+     */
+    suspend fun ensureSpeakerModel(progress: (String) -> Unit): SpeakerModelFile =
+        withContext(Dispatchers.IO) {
+            val state = speakerPath()
+            if (state.present) {
+                progress("speaker model present (${File(state.path).length() / 1_000_000}MB), skip")
+                return@withContext state
+            }
+            val task = DownloadTask(
+                File(state.path),
+                listOf(
+                    // hf-mirror works without login in China; HF直连为备份
+                    "https://hf-mirror.com/csukuangfj/speaker-embedding-models/resolve/main/3dspeaker_speech_campplus_sv_zh-cn_16k-common.onnx",
+                    "https://huggingface.co/csukuangfj/speaker-embedding-models/resolve/main/3dspeaker_speech_campplus_sv_zh-cn_16k-common.onnx",
+                ),
+                expectedSize = 28_281_138L,
+            )
+            downloadWithResume(http(), task, progress)
+            speakerPath()
+        }
+
+    private fun http() = OkHttpClient.Builder()
+        .connectTimeout(15, TimeUnit.SECONDS)
+        .readTimeout(120, TimeUnit.SECONDS)
+        .build()
+
     private data class DownloadTask(
         val target: File,
         val urls: List<String>,
+package com.lemon.lifelog.asr
+
+import android.util.Log
+import com.k2fsa.sherpa.onnx.SpeakerEmbeddingExtractor
+import com.k2fsa.sherpa.onnx.SpeakerEmbeddingExtractorConfig
+import java.io.File
+import java.nio.ByteBuffer
+import java.nio.ByteOrder
+import kotlin.math.sqrt
+
+/**
+ * On-device speaker embedding via sherpa-onnx (3dspeaker CAMPPlus, 192-dim).
+ *
+ * Use case (v0.3 MVP):
+ *   - For every VAD segment compute a 192-d float embedding.
+ *   - User enrolls 1+ "anchor profiles" (e.g. {name: "我", embedding}).
+ *   - Match new segments by cosine similarity; if max > threshold → assign that profile.
+ *
+ * Thread-unsafe per stream.  Reuse the extractor across segments.
+ */
+class SpeakerEmbedder private constructor(
+    private val extractor: SpeakerEmbeddingExtractor,
+    val dim: Int,
+) {
+
+    /** Compute embedding for a PCM segment. Returns null if not enough audio. */
+    fun embed(samples: FloatArray, sampleRate: Int = 16000): FloatArray? {
+        if (samples.isEmpty()) return null
+        val stream = extractor.createStream()
+        return try {
+            stream.acceptWaveform(samples, sampleRate)
+            stream.inputFinished()
+            if (!extractor.isReady(stream)) {
+                Log.w(TAG, "extractor not ready for ${samples.size} samples")
+                return null
+            }
+            extractor.compute(stream)
+        } finally {
+            stream.release()
+        }
+    }
+
+    fun release() {
+        extractor.release()
+    }
+
+    companion object {
+        private const val TAG = "SpeakerEmbedder"
+        const val DIM = 192
+
+        fun load(modelPath: String, numThreads: Int = 1): SpeakerEmbedder {
+            require(File(modelPath).exists()) { "speaker model not found: $modelPath" }
+            val t0 = System.nanoTime()
+            val extractor = SpeakerEmbeddingExtractor(
+                config = SpeakerEmbeddingExtractorConfig(
+                    model = modelPath,
+                    numThreads = numThreads,
+                    debug = false,
+                    provider = "cpu",
+                ),
+            )
+            val dim = extractor.dim()
+            Log.i(TAG, "loaded in ${(System.nanoTime() - t0) / 1_000_000}ms dim=$dim threads=$numThreads")
+            return SpeakerEmbedder(extractor, dim)
+        }
+    }
+}
+
+// -------- helpers --------------------------------------------------------
+
+/** Pack a FloatArray as little-endian bytes for Room BLOB column. */
+fun FloatArray.toLeBytes(): ByteArray {
+    val buf = ByteBuffer.allocate(size * 4).order(ByteOrder.LITTLE_ENDIAN)
+    for (f in this) buf.putFloat(f)
+    return buf.array()
+}
+
+/** Inverse of [toLeBytes]. */
+fun ByteArray.toFloatArrayLe(): FloatArray {
+    val n = size / 4
+    val out = FloatArray(n)
+    val buf = ByteBuffer.wrap(this).order(ByteOrder.LITTLE_ENDIAN).asFloatBuffer()
+    buf.get(out)
+    return out
+}
+
+/** Cosine similarity ∈ [-1,1]. Higher = more similar. */
+fun cosineSim(a: FloatArray, b: FloatArray): Float {
+    if (a.size != b.size || a.isEmpty()) return 0f
+    var dot = 0.0
+    var na = 0.0
+    var nb = 0.0
+    for (i in a.indices) {
+        dot += a[i] * b[i]
+        na += a[i] * a[i]
+        nb += b[i] * b[i]
+    }
+    if (na == 0.0 || nb == 0.0) return 0f
+    return (dot / (sqrt(na) * sqrt(nb))).toFloat()
+}
+
+/** Average a list of embeddings into one (per-dim mean), normalised. */
+fun averageEmbeddings(embeddings: List<FloatArray>): FloatArray {
+    require(embeddings.isNotEmpty()) { "empty list" }
+    val dim = embeddings.first().size
+    val sum = FloatArray(dim)
+    for (e in embeddings) {
+        require(e.size == dim) { "dim mismatch" }
+        for (i in 0 until dim) sum[i] += e[i]
+    }
+    val n = embeddings.size.toFloat()
+    for (i in 0 until dim) sum[i] /= n
+    return sum
+}
@@ -6,13 +6,21 @@ import androidx.room.Room
 import androidx.room.RoomDatabase
 @Database(
-    entities = [Segment::class, Transcript::class, Summary::class],
-    version = 1,
+    entities = [
+        Segment::class,
+        Transcript::class,
+        Summary::class,
+        SpeakerProfile::class,
+        TranscriptFts::class,
+    ],
+    version = 2,
+    exportSchema = false,
 )
 abstract class AppDatabase : RoomDatabase() {
     abstract fun segmentDao(): SegmentDao
     abstract fun transcriptDao(): TranscriptDao
     abstract fun summaryDao(): SummaryDao
+    abstract fun speakerDao(): SpeakerDao
     companion object {
         @Volatile
@@ -24,7 +32,10 @@ abstract class AppDatabase : RoomDatabase() {
                     ctx.applicationContext,
                     AppDatabase::class.java,
                     "lifelog.db",
-                ).build().also { instance = it }
+                )
+                    .fallbackToDestructiveMigration()  // v0.3 dev: schema 大改，旧测试数据可丢
+                    .build()
+                    .also { instance = it }
             }
     }
 }
@@ -11,6 +11,9 @@ data class SegmentRow(
     val durationMs: Long,
     val text: String?,
     val audioPath: String,
+    val speakerId: Long?,
+    val speakerSim: Float?,
+    val speakerName: String?,
 )
 @Dao
@@ -24,17 +27,52 @@ interface SegmentDao {
     @Query("SELECT * FROM segments ORDER BY startTs DESC")
     fun observeAll(): Flow<List<Segment>>
+    @Query("SELECT * FROM segments WHERE id = :id")
+    suspend fun get(id: Long): Segment?
+
+    @Query("UPDATE segments SET speakerId = :speakerId, speakerSim = :sim WHERE id = :id")
+    suspend fun updateSpeaker(id: Long, speakerId: Long?, sim: Float?)
+
+    /**
+     * Backfill: re-assign speakerId for all segments that have a stored embedding,
+     * using the latest profile set. Called once after a new profile is enrolled.
+     * (Actual cosine math is done in Kotlin — Room can't do it.)
+     */
+    @Query("SELECT * FROM segments WHERE embedding IS NOT NULL")
+    suspend fun allWithEmbedding(): List<Segment>
+
     @Query(
         """
         SELECT s.id AS id, s.startTs AS startTs, s.durationMs AS durationMs,
-               t.text AS text, s.audioPath AS audioPath
+               t.text AS text, s.audioPath AS audioPath,
+               s.speakerId AS speakerId, s.speakerSim AS speakerSim,
+               sp.name AS speakerName
         FROM segments s
         LEFT JOIN transcripts t ON t.segmentId = s.id
+        LEFT JOIN speaker_profiles sp ON sp.id = s.speakerId
         ORDER BY s.startTs DESC
         LIMIT :limit
         """,
     )
     fun observeTimeline(limit: Int = 100): Flow<List<SegmentRow>>
+
+    /** Full-text search via FTS4 virtual table. Falls back to LIKE if MATCH query is empty. */
+    @Query(
+        """
+        SELECT s.id AS id, s.startTs AS startTs, s.durationMs AS durationMs,
+               t.text AS text, s.audioPath AS audioPath,
+               s.speakerId AS speakerId, s.speakerSim AS speakerSim,
+               sp.name AS speakerName
+        FROM segments s
+        INNER JOIN transcripts t ON t.segmentId = s.id
+        INNER JOIN transcript_fts fts ON fts.docid = t.segmentId
+        LEFT JOIN speaker_profiles sp ON sp.id = s.speakerId
+        WHERE transcript_fts MATCH :query
+        ORDER BY s.startTs DESC
+        LIMIT :limit
+        """,
+    )
+    fun searchTimeline(query: String, limit: Int = 200): Flow<List<SegmentRow>>
 }
 @Dao
@@ -54,3 +92,18 @@ interface SummaryDao {
     @Query("SELECT * FROM summaries ORDER BY windowStartTs DESC LIMIT :n")
     fun recent(n: Int): Flow<List<Summary>>
 }
+
+@Dao
+interface SpeakerDao {
+    @Insert
+    suspend fun insert(p: SpeakerProfile): Long
+
+    @Query("SELECT * FROM speaker_profiles ORDER BY createdAt ASC")
+    fun observeAll(): Flow<List<SpeakerProfile>>
+
+    @Query("SELECT * FROM speaker_profiles ORDER BY createdAt ASC")
+    suspend fun all(): List<SpeakerProfile>
+
+    @Query("DELETE FROM speaker_profiles WHERE id = :id")
+    suspend fun delete(id: Long)
+}
 package com.lemon.lifelog.data
 import androidx.room.Entity
+import androidx.room.Fts4
 import androidx.room.PrimaryKey
 @Entity(tableName = "segments")
@@ -11,6 +12,12 @@ data class Segment(
     val audioPath: String,
     val durationMs: Long,
     val energyDb: Float = 0f,
+    /** 192-dim float embedding packed as little-endian bytes (192*4=768 B). null = 还没算 */
+    val embedding: ByteArray? = null,
+    /** 匹配到的 SpeakerProfile.id，null = 没匹配上 / 还没匹配 */
+    val speakerId: Long? = null,
+    /** 与 matched profile 的余弦相似度，null = 没匹配 */
+    val speakerSim: Float? = null,
 )
 @Entity(tableName = "transcripts")
@@ -18,7 +25,7 @@ data class Transcript(
     @PrimaryKey val segmentId: Long,
     val text: String,
     val lang: String = "zh",
-    val engine: String = "xfyun",
+    val engine: String = "paraformer-zh-on-device",
     val confidence: Float = 1f,
     val reviewedAt: Long? = null,
 )
@@ -32,3 +39,21 @@ data class Summary(
     val model: String,
     val createdAt: Long,
 )
+
+@Entity(tableName = "speaker_profiles")
+data class SpeakerProfile(
+    @PrimaryKey(autoGenerate = true) val id: Long = 0,
+    val name: String,
+    /** 192-dim float embedding packed as LE bytes，平均或单段 */
+    val embedding: ByteArray,
+    /** 用了多少段平均出来的 */
+    val sampleCount: Int = 1,
+    val createdAt: Long = System.currentTimeMillis(),
+)
+
+/** FTS4 全文索引，内容跟随 Transcript 表。Room 自动建表 + 触发器同步。 */
+@Fts4(contentEntity = Transcript::class, tokenizer = "unicode61")
+@Entity(tableName = "transcript_fts")
+data class TranscriptFts(
+    val text: String,
+)
@@ -7,4 +7,5 @@ class Repository(ctx: Context) {
     val segments = db.segmentDao()
     val transcripts = db.transcriptDao()
     val summaries = db.summaryDao()
+    val speakers = db.speakerDao()
 }
@@ -16,10 +16,15 @@ import com.k2fsa.sherpa.onnx.VadModelConfig
 import com.lemon.lifelog.MainActivity
 import com.lemon.lifelog.asr.ModelDownloader
 import com.lemon.lifelog.asr.ParaformerEngine
+import com.lemon.lifelog.asr.SpeakerEmbedder
+import com.lemon.lifelog.asr.cosineSim
+import com.lemon.lifelog.asr.toFloatArrayLe
+import com.lemon.lifelog.asr.toLeBytes
 import com.lemon.lifelog.audio.AudioCapture
 import com.lemon.lifelog.audio.SegmentWriter
 import com.lemon.lifelog.data.Repository
 import com.lemon.lifelog.data.Segment
+import com.lemon.lifelog.data.SpeakerProfile
 import com.lemon.lifelog.data.Transcript
 import java.io.File
 import kotlinx.coroutines.CoroutineScope
@@ -50,6 +55,15 @@ class RecorderService : Service() {
     private lateinit var segWriter: SegmentWriter
     private var vad: Vad? = null
     private var engine: ParaformerEngine? = null
+    private var speakerEmbedder: SpeakerEmbedder? = null
+
+    /** Anchor profiles cached in memory; refreshed when DB changes. */
+    @Volatile
+    private var profiles: List<CachedProfile> = emptyList()
+    private data class CachedProfile(val id: Long, val name: String, val embedding: FloatArray)
+
+    /** Cosine threshold above which we assign a profile to a segment. Tuned for CAMPPlus. */
+    private val speakerMatchThreshold = 0.5f
     override fun onCreate() {
         super.onCreate()
@@ -72,6 +86,7 @@ class RecorderService : Service() {
         capture.stop()
         vad?.release()
         engine?.release()
+        speakerEmbedder?.release()
         scope.cancel()
         wakeLock?.takeIf { it.isHeld }?.release()
         wakeLock = null
@@ -130,10 +145,30 @@ class RecorderService : Service() {
             ),
         )
+        // 2c. Speaker embedder (optional; download is 28 MB, may take a bit on cold start).
+        try {
+            RecorderState.setStatus("loading speaker model…")
+            val speakerFile = downloader.ensureSpeakerModel { msg ->
+                Log.i(TAG, "speaker dl: $msg")
+                RecorderState.setStatus(msg)
+            }
+            if (speakerFile.present) {
+                speakerEmbedder = SpeakerEmbedder.load(speakerFile.path, numThreads = 2)
+            }
+        } catch (t: Throwable) {
+            Log.w(TAG, "speaker embedder not available; segments will skip embedding", t)
+        }
+
+        // 2d. Load anchor profiles (cosine matching).
+        refreshProfiles()
+        scope.launch {
+            repo.speakers.observeAll().collect { _ -> refreshProfiles() }
+        }
+
         // 3. Start capturing.
         capture.start(scope)
         RecorderState.setStatus("listening")
-        Log.i(TAG, "pipeline running")
+        Log.i(TAG, "pipeline running speakers=${profiles.size}")
         capture.frames.collect { samples ->
             val v = vad ?: return@collect
@@ -153,12 +188,12 @@ class RecorderService : Service() {
         val now = System.currentTimeMillis()
         val startTs = now - audioMs
-        // a. persist audio (off CPU thread)
+        // a. persist audio (off CPU thread, Opus encode).
         val audioPath = kotlinx.coroutines.withContext(Dispatchers.IO) {
             segWriter.write(samples, startTs, sr)
         }
-        // b. paraformer
+        // b. paraformer.
         val wallStart = System.nanoTime()
         val text = try {
             engine!!.transcribeFloat(samples, sr).text
@@ -168,7 +203,33 @@ class RecorderService : Service() {
         }
         val wallMs = (System.nanoTime() - wallStart) / 1_000_000
-        // c. Room
+        // c. speaker embedding + nearest-profile match.
+        val embedding: FloatArray? = speakerEmbedder?.let { embedder ->
+            try {
+                embedder.embed(samples, sr)
+            } catch (t: Throwable) {
+                Log.w(TAG, "embed failed", t)
+                null
+            }
+        }
+        val (matchedSpeakerId: Long?, matchedSim: Float?, matchedName: String?) =
+            if (embedding != null && profiles.isNotEmpty()) {
+                var bestId: Long? = null
+                var bestName: String? = null
+                var bestSim = -1f
+                for (p in profiles) {
+                    val sim = cosineSim(embedding, p.embedding)
+                    if (sim > bestSim) {
+                        bestSim = sim
+                        bestId = p.id
+                        bestName = p.name
+                    }
+                }
+                if (bestSim >= speakerMatchThreshold) Triple(bestId, bestSim, bestName)
+                else Triple(null, bestSim, null)
+            } else Triple(null, null, null)
+
+        // d. Room insert.
         val segmentId = kotlinx.coroutines.withContext(Dispatchers.IO) {
             val id = repo.segments.insert(
                 Segment(
@@ -177,6 +238,9 @@ class RecorderService : Service() {
                     audioPath = audioPath,
                     durationMs = audioMs,
                     energyDb = 0f,
+                    embedding = embedding?.toLeBytes(),
+                    speakerId = matchedSpeakerId,
+                    speakerSim = matchedSim,
                 ),
             )
             if (text.isNotEmpty()) {
@@ -184,19 +248,29 @@ class RecorderService : Service() {
                     Transcript(
                         segmentId = id,
                         text = text,
-                        lang = "zh",
-                        engine = "paraformer-zh-on-device",
-                        confidence = 1f,
                     ),
                 )
             }
             id
         }
-        Log.i(TAG, "segment#$segmentId audio=${audioMs}ms wall=${wallMs}ms text=$text")
+        Log.i(
+            TAG,
+            "segment#$segmentId audio=${audioMs}ms wall=${wallMs}ms " +
+                "speaker=${matchedName ?: "—"} sim=${matchedSim?.let { "%.2f".format(it) } ?: "—"} " +
+                "text=$text",
+        )
         RecorderState.onSegment(text, audioMs, wallMs)
     }
+    private suspend fun refreshProfiles() {
+        val rows = kotlinx.coroutines.withContext(Dispatchers.IO) { repo.speakers.all() }
+        profiles = rows.map {
+            CachedProfile(it.id, it.name, it.embedding.toFloatArrayLe())
+        }
+        Log.i(TAG, "loaded ${profiles.size} speaker profiles")
+    }
+
     private fun startForegroundCompat() {
         val notification = buildNotification()
         if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.UPSIDE_DOWN_CAKE) {