出现未识别标签保留字幕bug修复：给兜底，与最相似声纹合并

2026-09-11 18:23:04 +08:00 · 2026-09-11 18:23:04 +08:00 · 7892e35d2d
parent 236037a0b0
commit 7892e35d2d
3 changed files with 159 additions and 44 deletions
--- a/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt
+++ b/app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt
@ -165,6 +165,11 @@ class MainActivity : AppCompatActivity() {
    private val speakerCheckThread = HandlerThread("speaker-window").also { it.start() }
    private val speakerCheckHandler by lazy { Handler(speakerCheckThread.looper) }
    private val speakerSwitchDrop = 0.30f   // cos(window, reference) below this => switch
+    // 最近邻兜底阈值(用户要求:未识别归到最相似的已识别人):
+    // endpoint 精确匹配不中→nearest≥0.35 归入;再不行才注册新人
+    private val verdictFallbackMin = 0.35f
+    // 滚动检测 fast 空间的兜底阈值(仅影响换行标签显示,endpoint 仍权威)
+    private val rollingFallbackMin = 0.30f
    // 滚动检测在 fast 空间查"这人是谁"仅用于**显示标签**,阈值放宽到 0.4;
    // endpoint 权威判定仍用准确模型 + speakerThreshold(0.5)。
    private val fastSearchThreshold = 0.4f
@ -176,6 +181,10 @@ class MainActivity : AppCompatActivity() {
    @Volatile private var rollingLowHits = 0
    // C(最新者胜):每次投递滚动检测都刷新序号,队列里被超额的旧任务直接跳过
    @Volatile private var rollingSeq = 0L
+    // 兜底残料:当前句子槽位最近一次滚动窗口声纹。整句判定因碎片算不出
+    // embedding 时复用它做归属(用户要求 2026-09-11)
+    @Volatile private var rollingEmbSid = -1L
+    @Volatile private var rollingEmb: FloatArray? = null
    @Volatile private var pendingSwitch: SpeakerSwitch? = null
    @Volatile private var windowRefEmb: FloatArray? = null
    @Volatile private var activeLabelForWindow: String? = null
@ -493,6 +502,8 @@ class MainActivity : AppCompatActivity() {
                                    win, sampleRateInHz
                                ) ?: return@post
                                if (rseq != rollingSeq) return@post
+                                rollingEmbSid = owner
+                                rollingEmb = emb
                                // 计算耗时内句子可能已结算:回写前二次核销
                                if (seq != checkSerial) return@post
                                val ref = windowRefEmb
@ -660,35 +671,7 @@ class MainActivity : AppCompatActivity() {

                        // 整句声纹判定移出录音线程(停摆根因):后台算完
                        // 回主线程 confirmSpeaker 修正归属
-                        speakerCheckHandler.post {
-                            val emb = SpeakerRecognition.computeEmbedding(
-                                snap, sampleRateInHz
-                            )
-                            // 镜像嵌入:② 用句尾 2s 切片(与滚动检测同视角),
-                            // 整句 embedding 会拉低窗口匹配度,导致换行时查不到人
-                            val fastEmb = if (emb != null &&
-                                SpeakerRecognition.fastReady()
-                            ) {
-                                val tailSlice = if (snap.size > 32000)
-                                    snap.copyOfRange(
-                                        snap.size - 32000, snap.size
-                                    ) else snap
-                                SpeakerRecognition.computeEmbeddingFast(
-                                    tailSlice, sampleRateInHz
-                                )
-                            } else null
-                            val label = if (emb == null) null else
-                                SpeakerRecognition.registry.search(
-                                    emb, speakerThreshold
-                                )
-                                    ?: SpeakerRecognition.nextAutoLabel().also {
-                                        SpeakerRecognition.registry.addAuto(it, emb)
-                                        if (fastEmb != null)
-                                            SpeakerRecognition.fastRegistry
-                                                .addAuto(it, fastEmb)
-                                    }
-                            uiHandler.post { confirmSpeaker(sid, label, emb, fastEmb) }
-                        }
+                        postSpeakerVerdict(sid, snap)
                        // 定格 + 送检立即走主线程队列,不等声纹
                        uiHandler.post {
                            val accepted = submitRefineTask(
@ -702,9 +685,67 @@ class MainActivity : AppCompatActivity() {
                }
            }
        }
+        // Stop 兜底(软 endpoint):最后一句未完的转写同样走定格→判定→精修,
+        // 否则丢半句 + 灰"未识别"行滞留(用户反馈 2026-09-11)
+        if (sentenceId != -1L) {
+            val raw = onlineRecognizer.getResult(stream).text
+            val disp =
+                if (raw.isNotBlank() && isInAllowedLanguages(raw)) raw else ""
+            onlineRecognizer.reset(stream)
+            val sid = sentenceId
+            sentenceId = -1L
+            windowRefEmb = null
+            if (disp.isNotBlank() && bufferTotal > 0) {
+                val snap = flattenSamples()
+                val dur = bufferTotal / sampleRateInHz.toFloat()
+                resetBuffer()
+                postSpeakerVerdict(sid, snap)
+                uiHandler.post {
+                    val accepted = submitRefineTask(
+                        sid, snap, dur, "STOP", null, disp, disp
+                    )
+                    finalizeSentence(sid, accepted)
+                }
+            } else {
+                resetBuffer()
+                uiHandler.post { finalizeSentence(sid, false) }
+            }
+        }
        stream.release()
    }

+    /**
+     * 整句声纹判定:后台算准确 embedding + fast 镜像(句尾 2s 切片,与滚动
+     * 检测同视角),注册后回主线程 confirmSpeaker 修正归属。
+     * 自然 endpoint / 强制断句 / Stop 软 endpoint 共用。
+     */
+    private fun postSpeakerVerdict(sid: Long, snap: FloatArray) {
+        speakerCheckHandler.post {
+            var emb = SpeakerRecognition.computeEmbedding(snap, sampleRateInHz)
+            if (emb == null && rollingEmbSid == sid) {
+                // 兜底①:句子太碎、整句 embedding 算不出 → 复用该槽滚动窗口残料
+                emb = rollingEmb
+            }
+            val fastEmb = if (emb != null && SpeakerRecognition.fastReady()) {
+                val tailSlice = if (snap.size > 32000)
+                    snap.copyOfRange(snap.size - 32000, snap.size) else snap
+                SpeakerRecognition.computeEmbeddingFast(tailSlice, sampleRateInHz)
+            } else null
+            val label = if (emb == null) null else
+                (SpeakerRecognition.registry.search(emb, speakerThreshold)
+                    // 兜底②:精确匹配不中 → 归入最相似的已注册身份
+                    ?: SpeakerRecognition.registry.nearest(
+                        emb, verdictFallbackMin
+                    ))
+                    ?: SpeakerRecognition.nextAutoLabel().also {
+                        SpeakerRecognition.registry.addAuto(it, emb)
+                        if (fastEmb != null)
+                            SpeakerRecognition.fastRegistry.addAuto(it, fastEmb)
+                    }
+            uiHandler.post { confirmSpeaker(sid, label, emb, fastEmb) }
+        }
+    }
+
    // Keep only zh/ja/en output when firstPassLanguage == "auto":
    // letters must be ASCII latin (en), kana (ja), or CJK hanzi (zh/ja);
    // everything else (hangul, cyrillic, thai, arabic, accented latin ...)
@ -820,9 +861,11 @@ class MainActivity : AppCompatActivity() {
    }

    private fun renameSpeaker(oldKey: String, name: String) {
-        // 真名不得长得像内部键/显示名,避免与 S{n}→未识别N 的映射冲突
-        if (autoLabelRegex.matches(name)) {
-            Toast.makeText(this, "名字不能是 S+数字 形式", Toast.LENGTH_SHORT).show()
+        // 真名不得长得像内部键/显示名,避免与 S{n}→已识别人N 的映射冲突
+        if (autoLabelRegex.matches(name) ||
+            Regex("^(已识别人?\\d+|未识别\\d*)$").matches(name)) {
+            Toast.makeText(this, "名字不能是 S数字/已识别人N/未识别N 形式",
+                Toast.LENGTH_SHORT).show()
            return
        }
        // 声纹库升级:S{n} 的嵌入转为真名注册(真名可多组嵌入并存,
@ -1246,14 +1289,14 @@ class MainActivity : AppCompatActivity() {
        subtitleRecycler.visibility = View.VISIBLE
    }

-    // 内部标签 S0/S1/…(注册键,可被命名为真名)对外显示为 未识别1/未识别2…;
-    // 已命名者直接显示真名。
+    // 内部标签 S0/S1/…(已注册、能区分但尚未起名的身份)对外显示"已识别人1/2…";
+    // 已命名者显示真名;label==null(完全无法归属)才显示灰色"未识别"。
    private val autoLabelRegex = Regex("^S(\\d+)$")

    private fun displayName(label: String?): String {
        if (label == null) return "未识别"
        val m = autoLabelRegex.matchEntire(label)
-        return if (m != null) "未识别${m.groupValues[1].toInt() + 1}" else label
+        return if (m != null) "已识别人${m.groupValues[1].toInt() + 1}" else label
    }

    private fun colorIdxFor(label: String): Int =
@ -1524,8 +1567,14 @@ class MainActivity : AppCompatActivity() {
    private fun rollingSpeakerSearch(emb: FloatArray): String? =
        if (SpeakerRecognition.fastReady())
            SpeakerRecognition.fastRegistry.search(emb, fastSearchThreshold)
+                ?: SpeakerRecognition.fastRegistry.nearest(
+                    emb, rollingFallbackMin
+                )
        else
            SpeakerRecognition.registry.search(emb, speakerThreshold)
+                ?: SpeakerRecognition.registry.nearest(
+                    emb, verdictFallbackMin
+                )

    // ③ 滚动检测查到了已知人、但当前行还没有归属(占位"未识别"):
    // 直接把活跃归属采纳为他,等 endpoint 权威判定给行冠名,不再多切一行
@ -1533,6 +1582,15 @@ class MainActivity : AppCompatActivity() {
        if (lastConfirmedLabel == null) {
            lastConfirmedLabel = label
            activeLabelForWindow = label
+            // 修"灰色未识别滞留"来源之一:滚动已匹配到人,顺带给当前灰色
+            // 占位行冠名(能并回上方同人行则并),不必等 endpoint 权威判定
+            subtitleLines.lastOrNull()
+                ?.takeIf { it.open && it.label == null }?.let { gray ->
+                    gray.label = label
+                    gray.colorIdx = colorIdxFor(label)
+                    if (!mergeLineIntoPrevIfSame(gray, label)) notifyLine(gray)
+                    Log.i(TAG, "adopt named gray line -> ${displayName(label)}")
+                }
        }
    }

@ -1695,15 +1753,21 @@ class MainActivity : AppCompatActivity() {
                    copyMinutesButton.isEnabled = true
                }
            } catch (t: Throwable) {
-                Log.e(TAG, "SUMMARIZE failed", t)
+                if (t is SummarizeEmptyException) Log.i(TAG, "SUMMARIZE empty: ${t.message}")
+                else Log.e(TAG, "SUMMARIZE failed", t)
                runOnUiThread {
                    minutesStatus.visibility = View.VISIBLE
-                    minutesStatus.text = if (t is SummarizeCancelledException)
-                        "已取消生成(保留上一次的纪要)"
-                    else "生成失败:${t.message}"
-                    Toast.makeText(this,
-                        if (t is SummarizeCancelledException) "已取消生成"
-                        else "纪要生成失败:${t.message}", Toast.LENGTH_LONG).show()
+                    minutesStatus.text = when (t) {
+                        is SummarizeCancelledException -> "已取消生成(保留上一次的纪要)"
+                        is SummarizeEmptyException ->
+                            "已自动取消:${t.message}(上次纪要未覆盖)"
+                        else -> "生成失败:${t.message}"
+                    }
+                    Toast.makeText(this, when (t) {
+                        is SummarizeCancelledException -> "已取消生成"
+                        is SummarizeEmptyException -> "已自动取消生成"
+                        else -> "纪要生成失败:${t.message}"
+                    }, Toast.LENGTH_LONG).show()
                }
            } finally {
                summarizeEngine = null
--- a/app/src/main/java/com/k2fsa/sherpa/onnx/SpeakerEmbedding.kt
+++ b/app/src/main/java/com/k2fsa/sherpa/onnx/SpeakerEmbedding.kt
@ -123,6 +123,26 @@ class SpeakerRegistry {
        return best
    }

+    /**
+     * 最近邻兜底(用户要求 2026-09-11):"未识别"归到与 e 最相似的已注册者
+     * (命名条目优先,其后自动条目;相似度须 ≥ minScore)。全空/不达标返回 null。
+     */
+    @Synchronized
+    fun nearest(e: FloatArray, minScore: Float): String? {
+        val q = l2normalize(e)
+        var bestName: String? = null
+        var best = -1f
+        for ((name, emb) in named) {
+            val s = dot(q, emb)
+            if (s > best) { best = s; bestName = name }
+        }
+        for ((name, emb) in auto) {
+            val s = dot(q, emb)
+            if (s > best) { best = s; bestName = name }
+        }
+        return if (best >= minScore) bestName else null
+    }
+
    /**
     * 点击标签命名用:把内部键 S{n}(自动注册、未命名)升级为真名,
     * 或把已有真名改成另一个真名。同名允许多条嵌入并存 —— 之前误判成
--- a/app/src/main/java/com/k2fsa/sherpa/onnx/SummarizePipeline.kt
+++ b/app/src/main/java/com/k2fsa/sherpa/onnx/SummarizePipeline.kt
@ -16,6 +16,9 @@ import org.json.JSONObject

 class SummarizeCancelledException : Exception("summarize cancelled")

+/** 空笔记防线(用户要求 2026-09-11):笔记/成稿为空 → 提示并自动取消,不落盘。 */
+class SummarizeEmptyException(msg: String) : Exception(msg)
+
 data class Prompts(
    val p1System: String, val p1UserTpl: String,
    val p2System: String, val p2UserTpl: String,
@ -167,6 +170,14 @@ class SummarizePipeline(
            else -> "English (formal written English)"
        }

+        /** 笔记是否无任何实质条目(四类条目全空;"前回末尾"不算)。解析失败按空处理。 */
+        internal fun notesEntriesEmpty(notes: String): Boolean = runCatching {
+            val o = JSONObject(notes)
+            listOf("已定事項", "討論中", "行動項目", "人物").all { k ->
+                o.optJSONArray(k)?.length() == 0
+            }
+        }.getOrDefault(true)
+
        fun mmss(ms: Long): String {
            val t = ms / 1000
            return "%02d:%02d".format(t / 60, t % 60)
@ -318,7 +329,7 @@ class SummarizePipeline(
    }

    /**
-     * @param labelOf 内部说话人键(S{n}/真名)→ 显示名(未识别N/真名)
+     * @param labelOf 内部说话人键(S{n}/真名)→ 显示名(已识别人N/真名/未识别)
     * @param entries SessionTranscript 快照(定格顺序,不可为空)
     * @param resumeNotes/resumeFromIndex M2:会中滚动笔记已覆盖前 resumeFromIndex 句,
     *        仅补跑尾部(进度、成稿复用同一份笔记)
@ -345,6 +356,13 @@ class SummarizePipeline(
        for (e in entries.drop(startIdx)) lines.add(renderLine(e))
        val chunks = chunk(lines)
        val total = chunks.size
+        // 用户要求(2026-09-11):没有新句可消化(再点生成纪要,断点已覆盖全部
+        // 转写)时不再空跑"成稿 0/0",直接提示并自动取消
+        if (total == 0) {
+            throw SummarizeEmptyException(if (resumeNotes != null)
+                "本场会议没有新增转写,纪要已是最新"
+            else "没有可分析的转写内容")
+        }
        Log.i(TAG, "SUMMARIZE: ${entries.size} utterances -> $total chunks, " +
                "lang=$lang, resumeIdx=$startIdx")

@ -375,7 +393,15 @@ class SummarizePipeline(
                )
                parsed = extractJson(out)
            }
-            if (parsed != null) notes = parsed else failed++
+            if (parsed != null) notes = parsed
+            else {
+                failed++
+                // 及时止损:第一块都解析不出来,笔记链条已断,后面块只会白烧 token
+                if (idx == 0) {
+                    Log.w(TAG, "P1 first chunk unparsable, abort. raw=${out.take(120)}")
+                    throw SummarizeEmptyException("开头转写片段无法生成有效笔记(内容过少或输出不可解析)")
+                }
+            }
            // 测量点(baseline,不改行为):raw≫json 说明解码浪费在废话/思考上;
            // ms 是单块真实耗时,决定 diff 化/大块合并的参数
            Log.i(TAG, "P1 chunk=${idx + 1}/$total prompt_len=${user.length} " +
@ -385,6 +411,10 @@ class SummarizePipeline(
        }
        Log.i(TAG, "P1 phase total_ms=${System.currentTimeMillis() - tP1Start} " +
                "chunks=$total failed=$failed")
+        // 空笔记终检:全部块失败,或笔记仍是空壳(total==0 已在入口取消)
+        if (failed == total || notesEntriesEmpty(notes)) {
+            throw SummarizeEmptyException("整场转写未能形成任何笔记条目")
+        }

        // ---- 成稿(§5.3)----
        if (listener.isCancelled()) throw SummarizeCancelledException()
@ -406,6 +436,7 @@ class SummarizePipeline(
        // 纯统计日志(零成本),纪要中的混语留给人工处理。
        val viol = scriptViolation(lang, md)
        if (viol != null) Log.i(TAG, "SUMMARIZE langcheck NOTE($lang): $viol (no rewrite)")
+        if (md.isBlank()) throw SummarizeEmptyException("成稿输出为空")
        Log.i(TAG, "SUMMARIZE wall clock total_ms=${System.currentTimeMillis() - tP1Start}")
        return MeetingSummary(md, lang, notes, total, failed)
    }