frontend: emit ㄭ for empty-rime syllables

scripts/frontend_bopomofo.py

@@ -14,7 +14,11 @@ ARPABET = ['AA','AE','AH','AO','AW','AY','B','CH','D','DH','EH','ER','EY','F','G
            'UH','UW','V','W','Y','Z','ZH']
 PUNCT = [',', '.', '?', '!', '…', '-', "'"]
 SPECIAL = ['_blank', '_pad', 'UNK', 'SP']            # SP = inter-word/space pause
-SYMBOLS = SPECIAL + ZHUYIN + ARPABET + PUNCT
+# ㄭ = syllabic-vowel symbol for empty-rime syllables (是/十/日/司...). U+312D is outside the
+# U+3105..U+3129 ZHUYIN range so it was missing. APPENDED AT END to preserve all existing phone ids
+# (warm-start compatibility); new id, embedding row trained during the re-align retrain.
+SYLLABIC = ['ㄭ']
+SYMBOLS = SPECIAL + ZHUYIN + ARPABET + PUNCT + SYLLABIC
 SYM2ID = {s: i for i, s in enumerate(SYMBOLS)}
 LANG = {'ZH': 0, 'EN': 1}                            # per-phone language id
 
@@ -34,6 +38,12 @@ def _split_syllable(syl: str):
     if syl and syl[-1].isdigit():
         tone = int(syl[-1]); syl = syl[:-1]
     units = [c for c in syl if c in SYM2ID]
+    # Empty-rime syllables (zhi/chi/shi/ri/zi/ci/si: 是/十/日/司/思/資...) are written in bopomofo as
+    # the bare retroflex/dental sibilant with an IMPLICIT syllabic vowel. Without an explicit vowel
+    # phone the model renders a clipped fricative that merges into the next syllable. Append the
+    # syllabic-vowel symbol ㄭ so these carry a proper rime. (Requires training data with ㄭ.)
+    if len(units) == 1 and units[0] in "ㄓㄔㄕㄖㄗㄘㄙ" and "ㄭ" in SYM2ID:
+        units = [units[0], "ㄭ"]
     return units, tone
 
 def text_to_phones(text: str):