feat: replace wanakana with english-to-kana 49K dictionary

- Vendor english-kana-matcher.js (english-to-kana, MIT, 49216 words) - Rewrite bridge: dictionary lookup + letter-by-letter fallback for OOV words - Convert English punctuation (,.!?) to Japanese pauses (、。) - Fix Chinese fullwidth comma being stripped by filter_kana - Preserve decimal points in numbers (3.5 -> スリーファイブ) - main.py now applies filter_kana (matches web UI behavior) - Remove wanakana dependency; rebuild portable zip (82.5 MB)
2026-08-08 15:06:34 +08:00
parent fd5c732d8a
commit 0c0e63161d
8 changed files with 152 additions and 40 deletions
@@ -1,17 +1,27 @@
 /** tts_bridge.js — Persistent Node.js bridge for aquestalk.js TTS synthesis.

-Converts English/romaji to katakana via wanakana before synthesis.
+Converts English words to katakana via vendored english-to-kana dictionary
+(english-kana-matcher.js, 49K words, MIT). Out-of-dictionary words fall back
+to letter-by-letter romanized spelling so no word is ever skipped.
+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";

 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);
+
+// ── 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 +34,6 @@ 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 +46,91 @@ try {
  process.exit(1);
 }

+// Letter-by-letter fallback readings (AquesTalk1-safe, no ヴ)
+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",
+};
+
+// English punctuation → Japanese pause equivalents
+const PUNCT_MAP = {
+  ",": "、", ".": "。", "!": "！", "?": "？",
+  "，": "、", "。": "。", "！": "！", "？": "？",
+};
+
+function kanaForChar(ch) {
+  return LETTER_KANA[ch.toLowerCase()] || DIGIT_KANA[ch] || null;
+}
+
+function wordToKana(word) {
+  const clean = word.replace(/'/g, "").toLowerCase();
+  const hit = lookupKana(clean);
+  if (hit) return hit;
+
+  // OOV fallback: spell out letter by letter (never skip)
+  let out = "";
+  for (const ch of clean) {
+    const k = kanaForChar(ch);
+    out += k ? k : "";
+  }
+  return out;
+}
+
+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(" "); // keep word gap for AquesTalk pause
+      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);
+  }
+
+  const wav = aq.run(result, speed);
  return Buffer.from(wav);
 }