重渲染机制更新，dsdur、dspitch提权

.gitignore

@@ -17,3 +17,4 @@ obj/
 # 打包产物
 *.tlx
 /build/
+/tools/CheckModel

DiffSingerDeclarations.cs

@@ -176,11 +176,11 @@ public static ObjectConfig BuildNoteConfig(VoicebankConfig config, INoteProperty
         DefaultOption = PropertyValue.Create(defaultValue),
     };
 
-    static List<ComboBoxOption> ToOptions(IReadOnlyList<string> values)
+    static List<ComboBoxOption> LanguageOptions(IReadOnlyList<string> languages)
     {
-        var options = new List<ComboBoxOption>(values.Count);
-        foreach (var value in values)
-            options.Add(value);   // 隐式转换：string → ComboBoxOption（值即显示文本）
+        var options = new List<ComboBoxOption> { new(PropertyValue.Create(string.Empty), "default") };
+        foreach (var lang in languages)
+            options.Add(lang);
         return options;
     }
 

DiffSingerForTuneLab.code-workspace

@@ -7,6 +7,10 @@
     {
       "name": "TuneLab (参考：SDK/docs/范例)",
       "path": "../TuneLab"
+    },
+    {
+      "name": "OpenUtau-lunai (参考：OpenUtau.Core)",
+      "path": "../OpenUtau-lunai"
     }
   ],
   "settings": {

DiffSingerModels.cs

@@ -121,6 +121,9 @@ public sealed class VoiceModels : IDisposable
     readonly Dictionary<string, DiffSingerPredictor?> mPredictors = new(StringComparer.Ordinal);
     readonly object mPredictorLock = new();
 
+    readonly object mAcousticLock = new();
+    readonly object mVocoderLock = new();
+
     public InferenceSession Acoustic { get; }
     public InferenceSession Vocoder { get; }
 

DiffSingerPhonemizer.cs

@@ -68,6 +68,35 @@ public static List<PhonemeSpan> Phonemize(
             string[] symbols = GetSymbols(dur, note, noteLang[i], out pinned[i]);
             noteSymbolCount[i] = symbols.Length;
 
+            // 连音符：不产生新音素，只延展前音的时长（通过 group 自然吸收）
+            // 连音符：- 不产生边界（前组自然吸收），+ 拆前组末音素到独立组强制 dur 边界
+            if (symbols.Length == 0 && (lyric == "-" || lyric == "+"))
+            {
+                if (lyric == "+")
+                {
+                    // 把前一个非空组的最后一个音素拆分到 + 组（仅当 >1 音素，否则退化为空组边界）
+                    string moved = "AP";
+                    bool splitted = false;
+                    for (int gi = groups.Count - 1; gi >= 0; gi--)
+                    {
+                        if (groups[gi].Phonemes.Count > 1)
+                        {
+                            int lastIdx = groups[gi].Phonemes.Count - 1;
+                            moved = groups[gi].Phonemes[lastIdx];
+                            groups[gi].Phonemes.RemoveAt(lastIdx);
+                            splitted = true;
+                            break;
+                        }
+                    }
+                    var g = new Group(note.StartTime, note.Pitch);
+                    if (splitted) g.Phonemes.Add(moved);
+                    groups.Add(g);
+                }
+                // - 则完全不做任何事，前组自然吸收时长，不影响 dur
+                notePhIndex.Add(notePhIndex[^1]);
+                continue;
+            }
+
             var wordGroups = ProcessWord(dur, note, symbols);
             groups[^1].Phonemes.AddRange(wordGroups[0].Phonemes);   // 前置辅音并入前一组（侵入前一 note 尾）
             groups.AddRange(wordGroups.Skip(1));                    // 韵核组（起点=note 起点）
@@ -169,17 +198,56 @@ static List<Group> ProcessWord(DiffSingerPredictor dur, SynthesisNoteSnapshot no
         return wordGroups;
     }
 
-    // 取音素符号串：钉死=用 note.Phonemes 符号；否则 G2P。过滤到「类型已定义 且 dur 表可 tokenize」；空则 [SP]。
+    // 取音素符号串：钉死/编辑器→用已有 phonemes 或 _phonemes 属性；连音符→空（slur 延展前音素）；否则 G2P。
+    // 空结果且非连音符→ [SP] 兜底。
     static string[] GetSymbols(DiffSingerPredictor dur, SynthesisNoteSnapshot note, string lang, out bool pinned)
     {
+        string lyric = note.Lyric ?? string.Empty;
+        if (lyric == "-" || lyric == "+")
+        {
+            pinned = false;
+            return Array.Empty<string>();
+        }
+
+        // 优先使用 _phonemes 属性（音素编辑器写入）
+        var phonemesProp = note.Properties.GetString("_phonemes", "");
+        if (!string.IsNullOrEmpty(phonemesProp) && phonemesProp != "[]")
+        {
+            pinned = true;
+            return ParsePhonemesProperty(phonemesProp).Where(s => !string.IsNullOrEmpty(s) && dur.TryPhoneme(s, out _)).ToArray();
+        }
+
+        // 其次使用钉死音素
         pinned = note.Phonemes.Count > 0;
         IEnumerable<string> raw = pinned
             ? note.Phonemes.Select(p => p.Symbol)
-            : dur.G2P(note.Lyric ?? string.Empty, lang);
+            : dur.G2P(lyric, lang);
         var symbols = raw.Where(s => dur.IsKnownSymbol(s) && dur.TryPhoneme(s, out _)).ToArray();
         return symbols.Length > 0 ? symbols : new[] { Pause };
     }
 
+    // 从 JSON 字符串解析音素符号列表：[{"s":"ja/b","v":false},...]
+    static string[] ParsePhonemesProperty(string json)
+    {
+        var result = new List<string>();
+        if (string.IsNullOrEmpty(json) || json.Length < 2) return result.ToArray();
+        try
+        {
+            int i = 0;
+            while (true)
+            {
+                int sIdx = json.IndexOf("\"s\":\"", i);
+                if (sIdx < 0) break;
+                sIdx += 5;
+                int eIdx = json.IndexOf('"', sIdx);
+                if (eIdx > sIdx) result.Add(json.Substring(sIdx, eIdx - sIdx));
+                i = eIdx + 1;
+            }
+        }
+        catch { }
+        return result.ToArray();
+    }
+
     // OpenUtau stretch：source[from..from+count) 的帧时长按 ratio 缩放、终点对齐 endPos，返回各音素起点秒。
     static IEnumerable<double> Stretch(IReadOnlyList<double> source, int from, int count, double ratio, double endPos)
     {
@@ -235,8 +303,7 @@ static double[] RunDur(DiffSingerPredictor dur, long[] tokens, long[] langs,
         var spk = new float[nTokens * hidden];
         for (int i = 0; i < nTokens; i++) Array.Copy(emb, 0, spk, i * hidden, hidden);
 
-        var durModel = dur.Model("dur");
-        var durInputs = new List<NamedOnnxValue>
+        using var durOut = dur.RunModel("dur", new List<NamedOnnxValue>
         {
             NamedOnnxValue.CreateFromTensor("encoder_out", encDense),
             NamedOnnxValue.CreateFromTensor("x_masks", maskDense),

DiffSingerPitch.cs

@@ -139,6 +139,7 @@ public static class DiffSingerPitch
             // 延音符（"-"/"+"）：无自身音素，沿用前一个 note 的 rest 状态（前为发声 ⇒ 本帧也发声、携自身 MIDI 滑过去）。
             if (DiffSingerPhonemizer.IsSlur(note.Lyric))
             {
+                // slur 延音符继承前音 rest 状态，确保 pitch 模型为其生成正确过渡音高
                 restList.Add(restList[^1]);
             }
             else

DiffSingerPredictor.cs

@@ -27,6 +27,8 @@ public sealed class DiffSingerPredictor : IDisposable
     readonly Dictionary<string, Dictionary<string, string[]>> mEntryCache = new(StringComparer.Ordinal);
     readonly Dictionary<string, string> mSymbolTypes = new(StringComparer.Ordinal);  // symbol → type（合并 dsdict）
     readonly object mLock = new();
+    // 推理锁：DirectML EP 的 InferenceSession.Run() 非线程安全，串行化所有 Run 调用。
+    readonly object mRunLock = new();
 
     public InferenceSession Linguistic { get; }
     public int HiddenSize => mHidden;
@@ -86,13 +88,21 @@ public int PhonemeToken(string symbol)
             : throw new InvalidOperationException($"音素 \"{symbol}\" 不在 {Path.GetFileName(mDir)} 的音素表中");
     public long LangId(string lang) => mLanguages.TryGetValue(lang, out var id) ? id : 0;
 
-    // —— G2P：按语言查 dsdict-{lang}.yaml 词条（grapheme→带前缀音素），exact 后小写回退 ——
+    // —— G2P：优先查语言特定词典（dsdict-{lang}.yaml），避免默认底库（dsdict.yaml 以 zh 为主）污染；再试 replacements；最后才兜底查合并词典。 ——
     public string[] G2P(string lyric, string lang)
     {
-        var entries = GetEntries(lang);
         var key = lyric.Trim();
-        if (entries.TryGetValue(key, out var phs)) return phs;
-        if (entries.TryGetValue(key.ToLowerInvariant(), out phs)) return phs;
+        // 1. 语言特定词典（不含默认底库）
+        var langEntries = GetLanguageSpecificEntries(lang);
+        if (langEntries.TryGetValue(key, out var phs)) return phs;
+        if (langEntries.TryGetValue(key.ToLowerInvariant(), out phs)) return phs;
+        // 2. 替换规则（en/ko 等无 entries 的语种）
+        var replaced = ApplyReplacements(lyric, lang);
+        if (replaced.Length > 0) return replaced;
+        // 3. 最后才查合并词典（含默认底库 dsdict.yaml，作为未知字素的最终兜底）
+        var allEntries = GetEntries(lang);
+        if (allEntries.TryGetValue(key, out phs)) return phs;
+        if (allEntries.TryGetValue(key.ToLowerInvariant(), out phs)) return phs;
         return Array.Empty<string>();
     }
 
@@ -119,7 +129,78 @@ public float[] GetEmbedding(string acousticSpeaker)
         }
     }
 
+    // —— 替换规则（用于 EN/KO 等无 entries 仅 replacements 的语种）——
+    readonly Dictionary<string, List<(string from, string to)>> mReplacements = new(StringComparer.Ordinal);
+
+    void LoadReplacements(string lang)
+    {
+        if (mReplacements.ContainsKey(lang)) return;
+        var list = new List<(string from, string to)>();
+        foreach (var file in new[] { $"dsdict-{lang}.yaml", $"dsdict-zh-{lang}.yaml", "dsdict.yaml" })
+        {
+            var path = Path.Combine(mDir, file);
+            if (!File.Exists(path)) continue;
+            try
+            {
+                var yaml = new DeserializerBuilder().Build();
+                var doc = yaml.Deserialize<Dictionary<string, object?>>(File.ReadAllText(path));
+                if (doc != null && doc.TryGetValue("replacements", out var reps) && reps is List<object?> repList)
+                {
+                    foreach (var r in repList)
+                    {
+                        if (r is Dictionary<object, object?> repDict)
+                        {
+                            string? from = repDict.TryGetValue("from", out var fv) ? fv?.ToString() : null;
+                            string? to = repDict.TryGetValue("to", out var tv) ? tv?.ToString() : null;
+                            if (!string.IsNullOrEmpty(from) && !string.IsNullOrEmpty(to))
+                                list.Add((from, to));
+                        }
+                    }
+                }
+            }
+            catch { }
+        }
+        mReplacements[lang] = list;
+    }
+
+    // 用替换规则将歌词转为音素（按最长匹配优先）
+    public string[] ApplyReplacements(string lyric, string lang)
+    {
+        LoadReplacements(lang);
+        if (!mReplacements.TryGetValue(lang, out var reps) || reps.Count == 0)
+            return Array.Empty<string>();
+
+        var repsSorted = reps.OrderByDescending(r => r.from.Length).ToList();
+        var result = new List<string>();
+        string text = lyric.ToLowerInvariant();
+        int pos = 0;
+        while (pos < text.Length)
+        {
+            bool matched = false;
+            foreach (var (from, to) in repsSorted)
+            {
+                if (pos + from.Length <= text.Length && text.Substring(pos, from.Length) == from)
+                {
+                    result.Add(to);
+                    pos += from.Length;
+                    matched = true;
+                    break;
+                }
+            }
+            if (!matched)
+            {
+                // 单个字符作为独立音素
+                string ch = text[pos].ToString();
+                result.Add(ch);
+                pos++;
+            }
+        }
+        return result.ToArray();
+    }
+
     // —— 词典加载 ——
+    // 策略：先加载 dsdict.yaml 作为默认底库，再叠加载入语种特定文件（后面覆盖前面）。
+    // 若 entries 为空且 replacements 存在，留空返回（上层调用 ApplyReplacements）。
     Dictionary<string, string[]> GetEntries(string lang)
     {
         lock (mLock)
@@ -128,21 +209,49 @@ Dictionary<string, string[]> GetEntries(string lang)
                 return cached;
 
             var map = new Dictionary<string, string[]>(StringComparer.Ordinal);
-            foreach (var file in new[] { $"dsdict-{lang}.yaml", $"dsdict-zh-{lang}.yaml", "dsdict.yaml" })
+
+            // 1. 加载默认底库 dsdict.yaml（总是存在）
+            var defaultPath = Path.Combine(mDir, "dsdict.yaml");
+            if (File.Exists(defaultPath))
+            {
+                var root = DeserializeDsDict(defaultPath);
+                foreach (var e in root.entries)
+                    if (!string.IsNullOrEmpty(e.grapheme))
+                        map[e.grapheme] = e.phonemes.ToArray();
+            }
+
+            // 2. 叠加载入语种特定文件（若存在则覆盖/补充）
+            foreach (var file in new[] { $"dsdict-{lang}.yaml", $"dsdict-zh-{lang}.yaml" })
             {
                 var path = Path.Combine(mDir, file);
                 if (!File.Exists(path)) continue;
                 var root = DeserializeDsDict(path);
                 foreach (var e in root.entries)
                     if (!string.IsNullOrEmpty(e.grapheme))
                         map[e.grapheme] = e.phonemes.ToArray();
-                break;
             }
+
             mEntryCache[lang] = map;
             return map;
         }
     }
 
+    // 仅加载语言特定词典（不含默认底库 dsdict.yaml），用于 G2P 的优先查表——避免 zh 底库污染其他语言的译音。
+    Dictionary<string, string[]> GetLanguageSpecificEntries(string lang)
+    {
+        var map = new Dictionary<string, string[]>(StringComparer.Ordinal);
+        foreach (var file in new[] { $"dsdict-{lang}.yaml", $"dsdict-zh-{lang}.yaml" })
+        {
+            var path = Path.Combine(mDir, file);
+            if (!File.Exists(path)) continue;
+            var root = DeserializeDsDict(path);
+            foreach (var e in root.entries)
+                if (!string.IsNullOrEmpty(e.grapheme))
+                    map[e.grapheme] = e.phonemes.ToArray();
+        }
+        return map;
+    }
+
     void LoadSymbolTypes(string dsdictPath)
     {
         if (!File.Exists(dsdictPath)) return;