fix: serialize TTSBridge IO and TTSService lifecycle to prevent concurrency crashes

- TTSBridge.synthesize wraps stdin/stdout request-response in asyncio.Lock;
  concurrent calls previously caused 'readuntil() called while another
  coroutine is already waiting' and ENOENT (temp dir deleted while bridge
  still reading input.txt)
- TTSService.start/stop guarded by lifecycle lock to prevent race from
  rapid Start clicks creating duplicate bridge/worker/danmaku connections
- Verified: 20 concurrent synthesize calls all succeed with correct WAV
- Rebuild portable zip (83.8 MB)

README.md

@@ -8,7 +8,7 @@ Bilibili 直播弹幕 + ゆっくり TTS 语音朗读器。
 
 - 实时接收 Bilibili 直播弹幕
 - 中文 → 拼音 → 片假名自动转换
-- 英文按罗马音转片假名（wanakana）
+- 英文转片假名（english-to-kana 49K 词库 + 字母拼读兜底）
 - 可选扫码登录（获取未打码用户名）
 - Web 控制面板（暗色主题）
 - 支持 8 种ゆっくり音色 + 语速/音量调节
@@ -94,7 +94,7 @@ Bilibili 弹幕 (blivedm)
 中文→片假名 (pypinyin + 映射表)
     │
     ▼
-英文→片假名 (wanakana)
+英文→片假名 (english-to-kana 49K 词库 + 字母拼读兜底)
     │
     ▼
 AquesTalk TTS (aquestalk.js + v86 WASM 模拟)
@@ -121,3 +121,48 @@ python build.py
 ## 许可
 
 MIT License
+
+```
+               .                                                                                 
+            -==- +                                                                                
+           ::.    %*%#*:                                                                            
+          --.. #***:*+********-.       .-+*#%%%%%%%##*+-:                          ..::-- -.        
+          .*---%#=.=*+*+*+******+#@%%#%%%#%%%#%#%#%%%%#%#%%#%@%-     .:-+###**+***+:+**:.-#+=       
+        ::+.-+++#+####****+**%#%%#%%#%%#%%#%#%%#%%#%#%%%%#%%#%#%%@@%#+**+****+**+**=-+#*::-+:       
+         *-.   #+.=+*+***%%%%%%#%%#%%#%#%#%#%#%%#%%#%#%#%#%#%%#%#%%%@@%%##*+***+**++*:=**#-.-+.     
+         *###**::*%%%#%#%#%#%#%%#%%#%%#%%#%%%#%#%#%%#%#%#%%#%#%%#%#%@@@@@%#*+*+**+**=-::-*#.  =     
+        +-#=%##%#%#%#%#%%%#%#%#%#%#%#%%#%%#%#%%%#%#%%%%#%%#%%%#%%#%#%#%@@@@@#########=+*+*+=:+.     
+       +---:%#%%%#%%#%%#%#%%%#%%#%%#%#%%#%#%%#%#%%%#%#%%#%%#%#%#%%#%#%%#%@%@@@######*++##--==       
+        =-##%%#%#%#%%#%%#%#%#%#%%#%%%#%#%%#%#%#%#%#%#%#%#%#%#%%%#%%%%#%#%#%@%@@%##**+++%#==-        
+        .%%%#%#%%%#%%#%#%%%#%%%%#%#%#%%#%#%%%#%%%#%%%#%%#%%#%#%#%@##%%#%%%%%%@%#+%##*+#:  :-:+      
+       :==#%#%%#%%%%#%%#%#%%#%@#%%%%%#%%#%%%%%#%#%#%#%#%%#%%%%%%%%@%%#%%##+-=-=-=-%#*++##**=:+      
+      ===--=--+%%%#%@%#%%#%%@@%%%#%@%%%%%#%@%#%#%%%#%@%%#%#%#%@%#%@**=--=--=+#%@@@%*+#%%=-%*        
+     :%#%#@%%%%==-----+*%%@@%@#%%%@%%@@#%%#@@%%%#%@%%#@%%%#*+-----=##@%#%@@@%%@%@%@%%%*-+:%@:       
+     %%#%@%%@@#%@@%%@%@%%@%@@%%%#%@@%@%%#%#@@@%%%%%@%@%@@%%%%@%@@@@%@@@%%%@@@%@@@@@%-+%##%@%*       
+    *%#%@@%@@#@@%@%@@@@%%@@@@%%%%@@@@#==%@%@%@@%%#%%@@@@@%%%%%%%@%@@@%@%%#%@@@@%@%@@#######%@.      
+   .%%%%@%@@%%@%@%%@%@%@@@%@%#@@%@%@#=-:*%@%@@%#%+%%%@%@%@%@@#%@@@%@@@@@%%%%%@@@@@%@####*+*+*%.     
+   =%#@@@@@%%@@@%@@@@+@@%@#-#%+@@@@#-=:..*@@@%@=%*=+%@@@@@@%@%%%@@@%@@%@@%%#%@%@%@@@###*+*####%*-*  
+   *%#@%@%@%@%@%@%@%-#%@%*+=##-@%@#==:....+@@@%-#*.:=@%@@=-%@@=*@%@@@%@@@@@%%%@@@@%@%##**+*--:++*-  
+   ##@@@%%#@@@@%@@*==*@====-%--+@*-=:... ..-%@*=.#%-=*#%%+=-+@*-=%@%@@@%@@%@@%%@%@@@%*+**#+-*%*+.   
+   *%@%@- @%@%@@@+==#+ =####%-....=-.........+-:.# :######%=*+%.-+@@@%@@@@@@%@@@@@%@%*####+=%%.     
+   =%%%. .@@@-#@=-=*:####*##%... ..... ... ......%####*##*%..*::.:@%@@*%%#%@@@%@%@@@#-:==+@@@@.     
+   :%#   -@--%%%--:- %******#........... .... ...*********+  +..+ *.=.-++-%@@@@@@@%@@%@%#%@@%@:     
+    :    :-#-==#:::-.=%#%%%%*.. ... .. .......... #######+ .:.:..=*=+#+#%==*%@%@%@@@@%@@@@@@@@:     
+           #***%-:---+.-................... .. .......---=+--=:-:-**+###%=*=*@@@@@%@@@@%@%@%@@:     
+          .#***%.:...:.......... .. .... ................:.:.:.:.-****###=-==%@%@@@%@%@@@@@@%@.     
+          .#*+##...:...:............. ...... .. .................-*+*####=-=-@@@%@@@@@@%@%@@@%.     
+           #***#............. .-... .........* ... ..............:**+###*=-=*%@@@%@@%@@@@@@@%%      
+          .#+**%.................#@**+========-..................:****##*=-+%@%@@@%@@@%@%@%@@%      
+         :=+=*##:......... ... ...*++========*..... .............:+-+*+#*=#@@@@@%@@@%@@@@@@*%*      
+          #.=*--*...... .... .... .:#====-==-#.. ..... .........=-*=#*#%+%@%@%@@@@%@@@@%@%@-@=      
+           -+ %@%-.. .................*+===*=....... .... .. ...=.+=-+#+=@%%@@%@@%@@@%@@@@*:%:      
+           :%%@%@@-..... .. ... ... ......... .. .................@%+@@@@%#%@%@#%@@@%=%@%@-:#       
+           :%%@@@%%#.. .......... ....... ........ ... .. ... . ==%#@%@%%%%%@@%-@@%@#:@@@% :*       
+           :%%%@%.%%@#:... .. ......... .... .. .... .........+- -%%%@@@%#%@%@:-@#-@+ %@%= =.       
+           :%%@@: %%-%@@*:..... ... ................... .. :+.   =#@@%=#*%%@@- =@:.@: #@#           
+           :%@@-  %#.-@@@@#%*:....... .. .. .... .. ...:=+.      #%@% -+-#@%=  *- .#. *%.           
+           +%@:   #:  @%@- *-  .=*+=-.... ... ...-++=.          .%@#  . .%@=   :   =  *.            
+           *@:    +   %#                                        =@-      %:                         
+           #.                                                  .*                                   
+
+```

build.py

@@ -135,13 +135,15 @@ def main() -> int:
     safe_copy_tree(aq_src / "node_modules", aq_dst / "node_modules",
                    ignore_patterns=("bililive-touhou-tts",))
 
-    # ── 4. kuroshiro node_modules ──────────────────────────────────────
-    step("Copying kuroshiro node_modules")
+    # ── 4. node_modules (if any) ───────────────────────────────────────
+    if (ROOT / "node_modules").exists():
+        step("Copying node_modules")
         safe_copy_tree(ROOT / "node_modules", BUILD_DIR / "node_modules")
 
     # ── 5. Bridge + config ─────────────────────────────────────────────
     step("Copying bridge files")
     shutil.copy2(ROOT / "tts_bridge.js", BUILD_DIR / "tts_bridge.js")
+    shutil.copy2(ROOT / "english-kana-matcher.js", BUILD_DIR / "english-kana-matcher.js")
     shutil.copy2(ROOT / "package.json", BUILD_DIR / "package.json")
     print("  OK")
 

chinese2kana.py

@@ -16,7 +16,7 @@ _KANA_SAFE_RE = re.compile(
     r"\u3001\u3002\uFF01\uFF1F"
     r"\u300C\u300D"
     r"\u30FB\u3000"
-    r"a-zA-Z0-9 ]"
+    r"a-zA-Z0-9 ,.!?\uFF0C]"
 )
 
 _CHINESE_DIGITS = ["零", "一", "二", "三", "四", "五", "六", "七", "八", "九"]
@@ -85,7 +85,7 @@ def chinese_to_kana(text: str, convert_numbers: bool = True) -> str:
         else:
             result_parts.append(seg)
     result = "".join(result_parts)
-    result = re.sub(r"(?<=\S) (?=\S)", "", result)
+    result = re.sub(r"(?<=[^\x00-\x7F]) (?=[^\x00-\x7F])", "", result)
     return result
 
 

main.py

@@ -12,7 +12,7 @@ import signal
 import sys
 
 from danmaku_handler import DanmakuClient
-from chinese2kana import chinese_to_kana
+from chinese2kana import chinese_to_kana, filter_kana
 from tts import TTSBridge
 from audio_player import play_wav_async, get_output_devices
 
@@ -47,6 +47,7 @@ async def tts_worker(queue: asyncio.Queue, bridge: TTSBridge,
             continue
         try:
             kana = chinese_to_kana(text, convert_numbers=convert_numbers)
+            kana = filter_kana(kana)
             logger.info("Speaking: %s -> %s", text, kana)
             wav_data = await bridge.synthesize(kana)
             await play_wav_async(wav_data)

package-lock.json

@@ -6,17 +6,30 @@
     "": {
       "name": "bililive-touhou-tts",
       "dependencies": {
-        "wanakana": "^5.3.1"
+        "phonemize": "^1.2.0"
       }
     },
-    "node_modules/wanakana": {
-      "version": "5.3.1",
-      "resolved": "https://registry.npmmirror.com/wanakana/-/wanakana-5.3.1.tgz",
-      "integrity": "sha512-OSDqupzTlzl2LGyqTdhcXcl6ezMiFhcUwLBP8YKaBIbMYW1wAwDvupw2T9G9oVaKT9RmaSpyTXjxddFPUcFFIw==",
+    "node_modules/number-to-words": {
+      "version": "1.2.4",
+      "resolved": "https://registry.npmmirror.com/number-to-words/-/number-to-words-1.2.4.tgz",
+      "integrity": "sha512-/fYevVkXRcyBiZDg6yzZbm0RuaD6i0qRfn8yr+6D0KgBMOndFPxuW10qCHpzs50nN8qKuv78k8MuotZhcVX6Pw==",
+      "license": "MIT"
+    },
+    "node_modules/phonemize": {
+      "version": "1.2.0",
+      "resolved": "https://registry.npmmirror.com/phonemize/-/phonemize-1.2.0.tgz",
+      "integrity": "sha512-+zEpOPXrEaylYCIXMSDVhiAwFbVzIXRf+7vheuNxozg4hLKbQVDXCOpI0GAJw41xEgE9LcwMjLUN30aWrpwSkw==",
       "license": "MIT",
-      "engines": {
-        "node": ">=12"
-      }
+      "dependencies": {
+        "number-to-words": "^1.2.4",
+        "pinyin-pro": "^3.26.0"
+      }
+    },
+    "node_modules/pinyin-pro": {
+      "version": "3.28.2",
+      "resolved": "https://registry.npmmirror.com/pinyin-pro/-/pinyin-pro-3.28.2.tgz",
+      "integrity": "sha512-jV38yxXHLfidirMC4hrXasLDozLCSq/4DfX88GnHcSEJ2+GpSedG6I9VOiEXJu6iQ5dbJC/RjmzyMuS5h/wH5A==",
+      "license": "MIT"
     }
   }
 }

package.json

@@ -2,6 +2,6 @@
   "name": "bililive-touhou-tts",
   "type": "module",
   "dependencies": {
-    "wanakana": "^5.3.1"
+    "phonemize": "^1.2.0"
   }
 }

server.py

@@ -37,6 +37,7 @@ class TTSService:
         self._shutdown_event: asyncio.Event | None = None
         self._worker_task: asyncio.Task | None = None
         self._cookies: dict[str, str] | None = None
+        self._lifecycle_lock = asyncio.Lock()
         self.running = False
         self.start_time: float = 0
         self.messages_processed = 0
@@ -48,6 +49,7 @@ class TTSService:
         self._cookies = cookies
 
     async def start(self, config: dict) -> None:
+        async with self._lifecycle_lock:
             if self.running:
                 raise RuntimeError("Already running")
             room_id = config["room_id"]
@@ -73,6 +75,7 @@ class TTSService:
             self._worker_task = asyncio.create_task(self._tts_worker(convert_numbers))
 
     async def stop(self) -> None:
+        async with self._lifecycle_lock:
             if not self.running:
                 return
             self.running = False

tts.py

@@ -45,6 +45,8 @@ class TTSBridge:
         self._process: asyncio.subprocess.Process | None = None
         self._bridge_script = bridge_script or _resolve_path("tts_bridge.js")
         self._node_exe = _get_node_exe()
+        self._stderr_task: asyncio.Task | None = None
+        self._io_lock = asyncio.Lock()
 
     async def start(self) -> None:
         if not pathlib.Path(self._bridge_script).exists():
@@ -61,6 +63,7 @@ class TTSBridge:
             stdout=asyncio.subprocess.PIPE,
             stderr=asyncio.subprocess.PIPE,
         )
+        self._stderr_task = asyncio.create_task(self._drain_stderr())
 
         line = await asyncio.wait_for(self._process.stdout.readline(), timeout=60)
         line_str = line.decode("utf-8").strip()
@@ -73,10 +76,28 @@ class TTSBridge:
 
         logger.info("Node.js TTS bridge ready.")
 
+    async def _drain_stderr(self) -> None:
+        """Continuously read bridge stderr and forward to Python logs."""
+        if self._process is None or self._process.stderr is None:
+            return
+        try:
+            while True:
+                raw = await self._process.stderr.readline()
+                if not raw:
+                    break
+                line = raw.decode("utf-8", errors="replace").rstrip()
+                if line:
+                    logger.info("[bridge] %s", line)
+        except asyncio.CancelledError:
+            pass
+        except Exception:
+            logger.debug("Bridge stderr drain stopped", exc_info=True)
+
     async def synthesize(self, kana_text: str) -> bytes:
         if self._process is None or self._process.stdin is None:
             raise RuntimeError("Bridge not started. Call start() first.")
 
+        async with self._io_lock:
             tmpdir = pathlib.Path(tempfile.mkdtemp(prefix="tts_"))
             try:
                 input_path = tmpdir / "input.txt"
@@ -114,6 +135,13 @@ class TTSBridge:
     async def stop(self) -> None:
         if self._process is not None:
             logger.info("Stopping TTS bridge...")
+            if self._stderr_task is not None:
+                self._stderr_task.cancel()
+                try:
+                    await self._stderr_task
+                except (asyncio.CancelledError, Exception):
+                    pass
+                self._stderr_task = None
             try:
                 if self._process.stdin is not None:
                     self._process.stdin.close()

tts_bridge.js

@@ -1,17 +1,40 @@
 /** tts_bridge.js — Persistent Node.js bridge for aquestalk.js TTS synthesis.
 
-Converts English/romaji to katakana via wanakana before synthesis.
+English word → katakana pipeline (never skips):
+  1. english-to-kana dictionary (49K words, MIT, vendored english-kana-matcher.js)
+  2. phonemize (MIT) G2P → IPA → rule-based katakana transcription
+  3. letter-by-letter romanized spelling (last resort)
+
+English punctuation (,.!?) is converted to Japanese pauses (、。).
 */
 
 import { readFileSync, writeFileSync } from "fs";
 import { createInterface } from "readline";
 import { fileURLToPath, pathToFileURL } from "url";
 import { dirname, join } from "path";
-import wanakana from "wanakana";
+import { createRequire } from "module";
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 
+// ── vendored english-to-kana dictionary (auto-generated, MIT) ──────────
+const matcherMod = pathToFileURL(join(__dirname, "english-kana-matcher.js")).href;
+const { lookupKana } = await import(matcherMod);
+
+// ── phonemize (G2P, MIT) via CJS entry (avoids JSON import issue in Node ESM) ──
+const require = createRequire(import.meta.url);
+let phonemize = null;
+try {
+  phonemize = require("phonemize").phonemize;
+  console.error("[bridge] phonemize G2P loaded.");
+} catch (err) {
+  console.error(`[bridge] phonemize unavailable (${err.message}), using letter fallback only.`);
+}
+
+// ── aquestalk.js ──────────────────────────────────────────────────────
+const aquestalkMod = pathToFileURL(join(__dirname, "aquestalk.js", "dist", "index.js")).href;
+const { load } = await import(aquestalkMod);
+
 const args = process.argv.slice(2);
 let voice = "f1";
 let speed = 100;
@@ -24,9 +47,6 @@ for (let i = 0; i < args.length; i++) {
   }
 }
 
-const aquestalkMod = pathToFileURL(join(__dirname, "aquestalk.js", "dist", "index.js")).href;
-const { load } = await import(aquestalkMod);
-
 console.error(`[bridge] Loading aquestalk.js with voice="${voice}", speed=${speed}...`);
 
 let aq;
@@ -39,11 +59,220 @@ try {
   process.exit(1);
 }
 
+// ── IPA → Katakana transcription (rule-based) ─────────────────────────
+const KANA = {
+  k: { a: "カ", i: "キ", u: "ク", e: "ケ", o: "コ" },
+  g: { a: "ガ", i: "ギ", u: "グ", e: "ゲ", o: "ゴ" },
+  s: { a: "サ", i: "シ", u: "ス", e: "セ", o: "ソ" },
+  z: { a: "ザ", i: "ジ", u: "ズ", e: "ゼ", o: "ゾ" },
+  t: { a: "タ", i: "チ", u: "トゥ", e: "テ", o: "ト" },
+  d: { a: "ダ", i: "ジ", u: "ドゥ", e: "デ", o: "ド" },
+  n: { a: "ナ", i: "ニ", u: "ヌ", e: "ネ", o: "ノ" },
+  h: { a: "ハ", i: "ヒ", u: "フ", e: "ヘ", o: "ホ" },
+  b: { a: "バ", i: "ビ", u: "ブ", e: "ベ", o: "ボ" },
+  p: { a: "パ", i: "ピ", u: "プ", e: "ペ", o: "ポ" },
+  m: { a: "マ", i: "ミ", u: "ム", e: "メ", o: "モ" },
+  r: { a: "ラ", i: "リ", u: "ル", e: "レ", o: "ロ" },
+  f: { a: "ファ", i: "フィ", u: "フ", e: "フェ", o: "フォ" },
+  v: { a: "バ", i: "ビ", u: "ブ", e: "ベ", o: "ボ" },
+  tʃ: { a: "チャ", i: "チ", u: "チュ", e: "チェ", o: "チョ" },
+  dʒ: { a: "ジャ", i: "ジ", u: "ジュ", e: "ジェ", o: "ジョ" },
+  ʃ: { a: "シャ", i: "シ", u: "シュ", e: "シェ", o: "ショ" },
+  ʒ: { a: "ジャ", i: "ジ", u: "ジュ", e: "ジェ", o: "ジョ" },
+  θ: { a: "サ", i: "シ", u: "ス", e: "セ", o: "ソ" },
+  ð: { a: "ザ", i: "ジ", u: "ズ", e: "ゼ", o: "ゾ" },
+  w: { a: "ワ", i: "ウィ", u: "ウ", e: "ウェ", o: "ウォ" },
+  j: { a: "ヤ", i: "イ", u: "ユ", e: "イェ", o: "ヨ" },
+  ŋ: { a: "ンガ", i: "ンギ", u: "ング", e: "ンゲ", o: "ンゴ" },
+};
+
+const CONS_ROW = {
+  "p": "p", "b": "b", "t": "t", "d": "d", "k": "k", "ɡ": "g", "g": "g",
+  "f": "f", "v": "v", "s": "s", "z": "z", "ʃ": "ʃ", "ʒ": "ʒ",
+  "h": "h", "tʃ": "tʃ", "dʒ": "dʒ", "θ": "θ", "ð": "ð",
+  "m": "m", "n": "n", "ŋ": "ŋ", "l": "r", "ɫ": "r", "ɹ": "r",
+  "r": "r", "j": "j", "w": "w", "ɾ": "r",
+};
+
+const VOWEL_KANA = { a: "ア", i: "イ", u: "ウ", e: "エ", o: "オ" };
+const DIPH_KANA = { ai: "アイ", au: "アウ", oi: "オイ" };
+
+function vClass(v) {
+  switch (v) {
+    case "ə": case "ɚ": case "ɝ": case "ɑ": case "ʌ": case "ɒ": case "æ": case "ɜ":
+      return { vowel: "a", long: false, diph: null };
+    case "ɪ": return { vowel: "i", long: false, diph: null };
+    case "i": return { vowel: "i", long: true, diph: null };
+    case "ʊ": return { vowel: "u", long: false, diph: null };
+    case "u": return { vowel: "u", long: true, diph: null };
+    case "ɛ": case "e": return { vowel: "e", long: false, diph: null };
+    case "eɪ": return { vowel: "e", long: true, diph: null };
+    case "ɔ": return { vowel: "o", long: false, diph: null };
+    case "o": case "oʊ": return { vowel: "o", long: true, diph: null };
+    case "aɪ": return { vowel: "a", long: false, diph: "ai" };
+    case "aʊ": return { vowel: "a", long: false, diph: "au" };
+    case "ɔɪ": return { vowel: "o", long: false, diph: "oi" };
+    default: return null;
+  }
+}
+
+const SONORANT = new Set(["m", "n", "ŋ", "l", "ɫ", "ɹ", "r", "j", "w", "ɾ"]);
+
+function ipaToKana(ipaStr) {
+  const tokens = [];
+  let i = 0;
+  while (i < ipaStr.length) {
+    const two = ipaStr.slice(i, i + 2);
+    if (CONS_ROW[two] || vClass(two)) { tokens.push(two); i += 2; continue; }
+    const one = ipaStr[i];
+    if (CONS_ROW[one] || vClass(one)) { tokens.push(one); i += 1; continue; }
+    if (one === " " || one === "ː") { tokens.push(one); i += 1; continue; }
+    i += 1;
+  }
+
+  let out = "";
+  for (let i = 0; i < tokens.length; i++) {
+    const t = tokens[i];
+    if (t === " ") { out += " "; continue; }
+    if (t === "ː") { out += "ー"; continue; }
+    const vc = vClass(t);
+    if (vc) {
+      out += vc.diph ? DIPH_KANA[vc.diph] : (VOWEL_KANA[vc.vowel] + (vc.long ? "ー" : ""));
+      continue;
+    }
+    const row = CONS_ROW[t];
+    if (!row) continue;
+
+    const next = tokens[i + 1];
+    const nextVc = next !== undefined ? vClass(next) : null;
+    if (nextVc) {
+      if (nextVc.diph) {
+        out += KANA[row][nextVc.vowel] + DIPH_KANA[nextVc.diph].slice(1);
+      } else {
+        out += KANA[row][nextVc.vowel] + (nextVc.long ? "ー" : "");
+      }
+      i++;
+    } else if (next !== undefined) {
+      out += KANA[row]["u"];
+    } else {
+      out += (t === "n" || t === "ŋ") ? "ン" : KANA[row]["u"];
+    }
+  }
+  return out;
+}
+
+// ── Letter-by-letter fallback (last resort, AquesTalk1-safe) ──────────
+const LETTER_KANA = {
+  a: "エー", b: "ビー", c: "シー", d: "ディー", e: "イー",
+  f: "エフ", g: "ジー", h: "エイチ", i: "アイ", j: "ジェー",
+  k: "ケー", l: "エル", m: "エム", n: "エヌ", o: "オー",
+  p: "ピー", q: "キュー", r: "アール", s: "エス", t: "ティー",
+  u: "ユー", v: "ブイ", w: "ダブリュー", x: "エックス", y: "ワイ", z: "ゼット",
+};
+
+const DIGIT_KANA = {
+  "0": "ゼロ", "1": "ワン", "2": "ツー", "3": "スリー", "4": "フォー",
+  "5": "ファイブ", "6": "シックス", "7": "セブン", "8": "エイト", "9": "ナイン",
+};
+
+function kanaForChar(ch) {
+  return LETTER_KANA[ch.toLowerCase()] || DIGIT_KANA[ch] || null;
+}
+
+function spellWord(word) {
+  let out = "";
+  for (const ch of word) {
+    const k = kanaForChar(ch);
+    out += k ? k : "";
+  }
+  return out;
+}
+
+// ── English punctuation → Japanese pause equivalents ──────────────────
+const PUNCT_MAP = {
+  ",": "、", ".": "。", "!": "！", "?": "？",
+  "，": "、", "。": "。", "！": "！", "？": "？",
+  " ": "、", // AquesTalk truncates on space; use 、 pause instead
+};
+
+function wordToKana(word) {
+  const clean = word.replace(/'/g, "").toLowerCase();
+  if (!clean) return "";
+
+  // 1. dictionary
+  const hit = lookupKana(clean);
+  if (hit) return hit;
+
+  // 2. phonemize G2P → IPA → katakana
+  if (phonemize && /^[a-z]+$/.test(clean)) {
+    try {
+      const ipa = phonemize(clean, { stripStress: true });
+      if (ipa && /[^a-z]/.test(ipa)) {
+        const kana = ipaToKana(ipa);
+        if (kana) return kana;
+      }
+    } catch (err) {
+      // fall through to letter spelling
+    }
+  }
+
+  // 3. letter-by-letter spelling (never skip)
+  return spellWord(clean);
+}
+
+function convertEnglishSegment(text) {
+  // Protect decimal points between digits (3.5) from being treated as periods
+  const protectedText = text.replace(/(\d)\.(\d)/g, "$1\u30FB$2");
+  const words = protectedText.split(/([\s]+|[.,!?，。！？]+)/).filter((s) => s.length > 0);
+  const parts = [];
+
+  for (const token of words) {
+    if (/^\s+$/.test(token)) {
+      parts.push("、"); // word gap: AquesTalk truncates on space, use 、 instead
+      continue;
+    }
+    if (PUNCT_MAP[token] !== undefined) {
+      parts.push(PUNCT_MAP[token]);
+      continue;
+    }
+    parts.push(wordToKana(token));
+  }
+
+  return parts.join("").replace(/、+/g, "、");
+}
+
 function synthesize(kanaText) {
   const trimmed = kanaText.trim();
   if (!trimmed) throw new Error("EMPTY_TEXT");
-  const finalText = wanakana.toKatakana(trimmed, { convertLongVowelMark: true });
-  const wav = aq.run(finalText, speed);
+
+  let result = "";
+  let englishBuf = "";
+
+  for (const ch of trimmed) {
+    if ((ch >= "a" && ch <= "z") || (ch >= "A" && ch <= "Z") ||
+        (ch >= "0" && ch <= "9") || ch === "'" || ch === "," || ch === "." ||
+        ch === "!" || ch === "?") {
+      englishBuf += ch;
+    } else {
+      if (englishBuf) {
+        result += convertEnglishSegment(englishBuf);
+        englishBuf = "";
+      }
+      // Full-width punctuation that slipped through (e.g. ，) → Japanese pause
+      result += PUNCT_MAP[ch] !== undefined ? PUNCT_MAP[ch] : ch;
+    }
+  }
+  if (englishBuf) {
+    result += convertEnglishSegment(englishBuf);
+  }
+
+  // Collapse consecutive pauses; strip any residual spaces (AquesTalk truncates on space)
+  result = result.replace(/ +/g, "、").replace(/、+/g, "、").replace(/^、|、$/g, "");
+
+  console.error(`[bridge] SYNTH IN:  ${trimmed}`);
+  console.error(`[bridge] SYNTH OUT: ${result}`);
+
+  const wav = aq.run(result, speed);
   return Buffer.from(wav);
 }