chore: auto-commit from dashboard

src/voice/normalize.py

@@ -94,6 +94,31 @@ def expand_numbers_ro(text: str) -> str:
     return _NUM_TOKEN.sub(_sub, text)
 
 
+def _decimal_to_en(s: str) -> str:
+    """Convert decimal string 'X.Y' to English words ('3.14' -> 'three point one four')."""
+    int_part, dec_part = s.split('.', 1)
+    int_words = num2words(int(int_part), lang='en')
+    dec_words = ' '.join(num2words(int(d), lang='en') for d in dec_part)
+    return f"{int_words} point {dec_words}"
+
+
+def expand_numbers_en(text: str) -> str:
+    """Expand bare numeric tokens to English words.
+
+    Mirrors expand_numbers_ro for the pocket-tts (English-only) path.
+    Must run after normalize_thousands so Romanian-style grouped
+    integers ("105.300") read as one magnitude ("one hundred five
+    thousand three hundred") instead of a decimal ("105.3").
+    """
+    def _sub(match: re.Match) -> str:
+        token = match.group(1)
+        if '.' in token:
+            return _decimal_to_en(token)
+        return num2words(int(token), lang='en')
+
+    return _NUM_TOKEN.sub(_sub, text)
+
+
 # ---------- Thousands separator ----------
 
 # Romanian uses dot or space as thousands separator: 384.000 / 384 000. The
@@ -268,8 +293,8 @@ def expand_currency(text: str) -> str:
 # ---------- Symbols ----------
 
 _SYMBOL_WORDS = {
-    'ro': {'%': ' la sută', '&': ' și ', '@': ' la ', '°': ' grade'},
-    'en': {'%': ' percent', '&': ' and ', '@': ' at ', '°': ' degrees'},
+    'ro': {'%': ' la sută', '&': ' și ', '@': ' la ', '°': ' grade', '~': ' aproximativ '},
+    'en': {'%': ' percent', '&': ' and ', '@': ' at ', '°': ' degrees', '~': ' about '},
 }
 
 
@@ -322,11 +347,14 @@ def expand_for_tts(text: str, lang: str = 'ro') -> str:
     "Restul l-am scris în chat." suffix from normalize_for_tts() would be
     misleading (no live chat mirror exists for that flow).
 
-    The RO-specific expansions (abbreviations, thousands, time, currency,
-    units, numbers-to-words) only make sense for Romanian text — for other
-    languages (e.g. English text routed to pocket-tts, which is
-    English-only) they'd inject Romanian words/diacritics into text the
-    target engine can't speak. Skip them when lang != 'ro'.
+    The RO-specific expansions (abbreviations, time, currency wording,
+    units) only make sense for Romanian text — for other languages (e.g.
+    English text routed to pocket-tts, which is English-only) they'd
+    inject Romanian words/diacritics into text the target engine can't
+    speak, so they're skipped when lang != 'ro'. Thousands-grouping and
+    numbers-to-words still run for lang == 'en', with English wording,
+    since replies on English turns often carry Romanian-formatted figures
+    ("105.300 lei") that pocket-tts would otherwise misread as decimals.
     """
     text = strip_markdown(text)
     text = sanitize_punctuation(text)
@@ -337,6 +365,16 @@ def expand_for_tts(text: str, lang: str = 'ro') -> str:
         text = expand_currency(text)
         text = expand_units(text)
         text = expand_numbers_ro(text)
+    elif lang == 'en':
+        # RO-specific expansions (abbreviations, time, currency wording,
+        # units) are skipped here — see docstring above. But bare numbers
+        # still need expanding: replies to Marius often carry Romanian
+        # thousands-grouped figures ("105.300 lei") even on English turns,
+        # and pocket-tts reads the dot as an English decimal point if left
+        # alone ("one hundred five point three" instead of "one hundred
+        # five thousand three hundred").
+        text = normalize_thousands(text)
+        text = expand_numbers_en(text)
     text = expand_symbols(text, lang=lang)
     return text.strip()