Feature/malayalam normalizer

omnivoice/models/omnivoice.py

@@ -69,6 +69,7 @@
 )
 from omnivoice.utils.duration import RuleDurationEstimator
 from omnivoice.utils.lang_map import LANG_IDS, LANG_NAMES
+from omnivoice.utils.text_normalization import normalize_text as _normalize_text_ml
 from omnivoice.utils.text import (
     add_punctuation,
     chunk_text_punctuation,
@@ -1061,6 +1062,16 @@ def _preprocess_all(
             text_list = [
                 _normalize_text(t, lang) for t, lang in zip(text_list, language_list)
             ]
+
+        # Apply language-specific text normalization (e.g., numbers, currency,
+        # units) so the TTS model receives spoken-form text instead of raw
+        # digits/symbols. Currently supports Malayalam ("ml").
+        for i in range(batch_size):
+            if language_list[i] is not None:
+                text_list[i] = _normalize_text_ml(
+                    text_list[i], language=language_list[i]
+                )
+
         instruct_list = self._ensure_list(instruct, batch_size)
         for i, s in enumerate(instruct_list):
             if s is None:

omnivoice/utils/text_normalization/__init__.py

@@ -0,0 +1,44 @@
+"""Text normalization utilities for TTS preprocessing.
+
+Provides language-aware text normalization that converts numbers, currency,
+percentages, and units into spoken-form text. This improves TTS pronunciation
+quality for non-Latin scripts where raw digits produce poor speech output.
+
+Currently supported languages:
+- Malayalam (``ml``)
+
+Usage::
+
+    from omnivoice.utils.text_normalization import normalize_text
+
+    text = normalize_text("എനിക്ക് ₹100 ഉണ്ട്", language="ml")
+    # Returns: "എനിക്ക് നൂറ് രൂപ ഉണ്ട്"
+"""
+
+from typing import Optional
+
+
+def normalize_text(text: str, language: Optional[str] = None) -> str:
+    """Normalize text for TTS based on the target language.
+
+    Converts numbers, currency symbols, percentages, and measurement units
+    into their spoken-word equivalents in the target language.
+
+    Args:
+        text: Input text that may contain raw numbers, currency, etc.
+        language: ISO 639 language code (e.g., ``"ml"`` for Malayalam).
+            If ``None`` or unsupported, the text is returned unchanged.
+
+    Returns:
+        Normalized text with numbers/symbols replaced by words.
+    """
+    if language is None or not text:
+        return text
+
+    if language == "ml":
+        from omnivoice.utils.text_normalization.malayalam import normalize_malayalam
+
+        return normalize_malayalam(text)
+
+    # Unsupported languages pass through unchanged
+    return text