乐观快切(用户拍板:第一次识别先切、出错由并回自愈):

- 开口即查去掉'等准确复核'挡切换:fast 偏离即切+占号,复核降级为善后
- 新增快速滚动通道:2s 尾窗/1s 周期,快模型偏离锚点 0.32 即乐观快切(无缝场景检测延迟 3~5s→1.5~2.5s 目标)
- accurateRecheckAfterSwitch 善后:认出当前人→留痕(判定链并回自愈);真新人→补写准确空间防重发号;认出别人→留痕(判定链迁回)
- 冷却/锚点自弃/滞后喂料/标签钉住全套保留;fastRollDrop 为灵敏度钮

app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt

@@ -221,6 +221,13 @@ class MainActivity : AppCompatActivity() {
     private val onsetMinSamples = 19200     // 1.2s 开口窗
     // 喂料恒定滞后(2026-09-16):识别流滞后 1.2s 喂入,切点前的音频可改道;
     // 全局字幕延迟 ~1.2s 的代价换取换人零串行
+    // 快速滚动通道(2026-09-16 用户拍板乐观快切):2s 尾窗/1s 周期/偏离
+    // 0.32 即切;准确复核只善后,切错由判定链并回自愈
+    private val fastRollEnabled = true
+    private val fastRollDrop = 0.32f
+    private val fastStepSamples = 16000     // 1s
+    private val fastWinSamples = 32000      // 2s 尾窗
+    private val fastMinSamples = 24000      // ≥1.5s 才判
     private val feedLagSamples = 24000   // 1.5s:覆盖判定链(1.2s采集+计算)的
     // 余量,开口渗漏≈0;全局字幕延迟即此值,调参范围 1.2~1.5s
     @Volatile private var onsetSeq = 0L     // 行切换/会话开始作废在途开口判定
@@ -636,6 +643,7 @@ class MainActivity : AppCompatActivity() {
         var newSinceCheck = 0
         var silenceRun = 0
         var onsetRun = 0   // 开口即查:静音后已累积的开口语音采样数
+        var fastSince = 0   // 快速滚动通道:1s 周期的采样计数
         // 恒定滞后喂料(2026-09-16):音频先进 pendingBuf,滞后 feedLagSamples
         // 再喂识别流/样本缓冲(窗口分析保持实时)。换人决定时冻结缓冲,
         // 断句后补喂给新流——新人开头音频从不进入旧句子。
@@ -688,48 +696,25 @@ class MainActivity : AppCompatActivity() {
                                 speakerCheckHandler.post {
                                     if (rseq != onsetSeq) return@post
                                     val anchor = windowRefEmb ?: return@post
-                                    // 一级:快锚点相似度(灵敏);贴近锚点=同人,免准确复核
                                     val fastEmb = SpeakerRecognition.computeEmbeddingFast(
                                         slice, sampleRateInHz) ?: return@post
                                     if (rseq != onsetSeq) return@post
                                     val simFast = SpeakerRecognition.similarity(
                                         fastEmb, anchor)
                                     if (simFast >= emaAbsorbMin) return@post
-                                    // 二级:准确空间复核(2026-09-16 误报实锤:A 暂停
+                                    // 乐观快切(2026-09-16 用户拍板:第一次识别
-                                    // 续说,fast 短窗 simA<0.5 误判新人→A 的续说句被切
+                                    // 先切,出错由判定链并回自愈;准确复核不再
-                                    // 成"已识别人2"首句)。同人暂停后短窗声纹弱,fast
+                                    // 挡在切换前,只做善后)
-                                    // 不可信;CAM++ 准确空间同人 0.77~0.93 强分离:
-                                    // 注册表认得出(且是当前人)→ 撤销切换;认得出别人
-                                    // → 切给他;谁都认不出 → 才当新人
-                                    val emb = SpeakerRecognition.computeEmbedding(
-                                        slice, sampleRateInHz) ?: return@post
-                                    if (rseq != onsetSeq) return@post
                                     val active = activeLabelForWindow
-                                    val known = SpeakerRecognition.registry.search(
+                                    val label = SpeakerRecognition.nextAutoLabel().also {
-                                        emb, speakerThreshold)
+                                        SpeakerRecognition.fastRegistry.addAuto(it, fastEmb)
-                                        ?: nearestAttribution(
-                                            emb, SpeakerRecognition.registry,
-                                            verdictFallbackMin)
-                                    if (known != null && known == active) {
-                                        // 复核撤销(同人暂停续说):留痕便于验证
-                                        Log.i(TAG, "ONSET-CANCEL fastSim=" +
-                                                "%.2f".format(simFast) +
-                                                " label=$known(active)")
-                                        return@post
                                     }
-                                    val label = known
-                                        ?: SpeakerRecognition.nextAutoLabel().also {
-                                            SpeakerRecognition.registry.addAuto(it, emb)
-                                            SpeakerRecognition.fastRegistry.addAuto(
-                                                it, fastEmb)
-                                        }
                                     Log.i(TAG, "ONSET-SWITCH fastSim=" +
                                             "%.2f".format(simFast) +
-                                            " label=$label kind=" +
+                                            " label=$label active=${active ?: "-"}")
-                                            (if (known != null) "known" else "new") +
-                                            " active=${active ?: "-"}")
                                     cutRequested = true
                                     uiHandler.post { onOnsetSwitch(label) }
+                                    accurateRecheckAfterSwitch(slice, label, active, "ONSET")
                                 }
                             }
                         }
@@ -738,6 +723,51 @@ class MainActivity : AppCompatActivity() {
                     }
                 }
 
+                // 快速滚动通道(2026-09-16 用户拍板:第一次识别先切、错由
+                // 并回自愈):2s 尾窗/1s 周期;快模型偏离锚点即乐观快切;
+                // 准确复核只善后(注册/留痕);3s 冷却与强切通道共用
+                if (fastRollEnabled && enableRollingSwitch && speakersEnabled &&
+                    !frozen
+                ) {
+                    fastSince += ret
+                    if (fastSince >= fastStepSamples) {
+                        fastSince = 0
+                        val winF = window.snapshot()
+                        val nF = fastWinSamples.coerceAtMost(winF.size)
+                        if (nF >= fastMinSamples && rmsOf(winF) >= minWindowRms) {
+                            val tailF = winF.copyOfRange(winF.size - nF, winF.size)
+                            val rseq = onsetSeq
+                            speakerCheckHandler.post {
+                                if (rseq != onsetSeq) return@post
+                                val anchor = windowRefEmb ?: return@post
+                                val emb = SpeakerRecognition.computeEmbeddingFast(
+                                    tailF, sampleRateInHz) ?: return@post
+                                if (rseq != onsetSeq) return@post
+                                val sim = SpeakerRecognition.similarity(emb, anchor)
+                                if (sim >= fastRollDrop) return@post
+                                val nowMs = SystemClock.elapsedRealtime()
+                                if (nowMs - lastForceSwitchAtMs < 3_000L) {
+                                    Log.i(TAG, "FASTROLL suppressed (cooldown)")
+                                    return@post
+                                }
+                                lastForceSwitchAtMs = nowMs
+                                val active = activeLabelForWindow
+                                val label = SpeakerRecognition.nextAutoLabel().also {
+                                    SpeakerRecognition.fastRegistry.addAuto(it, emb)
+                                }
+                                Log.i(TAG, "FASTROLL-SWITCH sim=" +
+                                        "%.2f".format(sim) +
+                                        " label=$label active=${active ?: "-"}")
+                                cutRequested = true
+                                uiHandler.post { onOnsetSwitch(label) }
+                                accurateRecheckAfterSwitch(tailF, label, active, "FASTROLL")
+                            }
+                        }
+                    }
+                } else {
+                    fastSince = 0
+                }
+
                 // P2(2026-09-11):跨句常驻检测,参考 embedding 为"行级"
                 // (appendNewLine 时重置)——新人开口后第一个周期内即可触发换行,
                 // 句间不再依赖 endpoint 的迟到判定
@@ -1759,6 +1789,26 @@ class MainActivity : AppCompatActivity() {
         Log.i(TAG, "ONSET-CLOSE no-sentence forced endpoint")
     }
 
+    /** 乐观切换后的准确空间善后(2026-09-16 用户拍板:先切后改):切换已
+     *  发出,本函数只做善后——认得出是当前人 → 留痕(切错由判定链"并回"
+     *  自愈);真新人 → 补写准确空间(防后续重发号);认出别人 → 留痕(判定
+     *  链迁回)。在 speakerCheckHandler 线程执行。 */
+    private fun accurateRecheckAfterSwitch(
+        samples: FloatArray, label: String, active: String?, tag: String,
+    ) {
+        val acc = SpeakerRecognition.computeEmbedding(samples, sampleRateInHz)
+            ?: return
+        val pre = SpeakerRecognition.registry.search(acc, speakerThreshold)
+            ?: nearestAttribution(acc, SpeakerRecognition.registry, verdictFallbackMin)
+        when {
+            pre != null && pre == active ->
+                Log.i(TAG, tag + "-CANCEL-LATE known=$pre(active) revert-via-verdict")
+            pre != null ->
+                Log.i(TAG, tag + "-RECHECK known=$pre label=$label (verdict migrates)")
+            else -> SpeakerRecognition.registry.addAuto(label, acc)
+        }
+    }
+
     private fun appendNewLine(label: String?): SubtitleLine {
         // 新行出现即封掉之前的开放行,并把其中"未定格的进行中句子"整体迁移
         // 过来 —— 硬保证 in-progress 只可能发生在最底行(修复:声纹判定搬家后,