Feature/malayalam normalizer

[spigelspike]

What this does

Adds text normalization for Malayalam (language="ml"). Numbers, currency symbols, time, percentages, etc. get converted to spoken Malayalam words before the text hits the tokenizer.

Without this, the model tries to pronounce raw symbols like ₹, %, 5:30 which sounds broken.

Examples

Input	Normalized
₹250	ഇരുനൂറ്റി അമ്പത് രൂപ
5:30	അഞ്ച് മുപ്പത്
AICTE	എ ഐ സി ടി ഇ
10.5	പത്ത് പോയിന്റ് അഞ്ച്
1/2	അര
1st	ഒന്നാമത്തെ
$100	നൂറ് ഡോളർ
10km	പത്ത് കിലോമീറ്റർ

Files changed

• omnivoice/utils/text_normalization/__init__.py — language dispatcher
• omnivoice/utils/text_normalization/malayalam.py — the normalizer
• omnivoice/models/omnivoice.py — calls normalize_text() inside _preprocess_all(), guarded by language check so other languages are unaffected
• tests/test_malayalam_normalizer.py — 58 tests
• docs/tips.md — added a note about this feature

Audio samples (before/after)

Test 1: "എന്റെ കയ്യിൽ $100 ഉം Rs. 50 ഉം ഉണ്ട്. ഇതിന് £200 വിലവരും."

• Baseline: baseline_sample1.wav
• With normalizer: custom_normalizer_sample1.wav

Test 2: "എനിക്ക് ₹100 ഉണ്ട്, സമയം 5:30 ആയി. AICTE മോഡൽ 10.5 സ്കോർ നേടി. 1/2 ലിറ്റർ പാലും 1st പ്രൈസും കിട്ടി"

• Baseline: baseline_sample2.wav
• With normalizer: custom_normalizer_sample2.wav

[spigelspike]

Added automatic Malayalam text normalization documentation. When language="ml" is specified, numbers, currency, percentages, and measurement units ETC are automatically normalized to improve TTS pronunciation.

[spigelspike]

Conflict resolved ✅

Rebased and merged with the latest master, including the recently merged #227 (opt-in text normalization via normalize_text=True).

How the two features coexist:

• Add opt-in text normalization (numbers, dates, currency) via generate(normalize_text=True) #227 adds general-purpose, opt-in normalization (normalize_text=True) for Chinese/English via WeTextProcessing + num2words fallback.
• This PR adds automatic, language-specific normalization for Malayalam (language="ml"). It runs unconditionally when the language is set to Malayalam, because Malayalam TTS quality degrades significantly without it (raw digits/symbols produce unintelligible speech).

Both pipelines are applied sequentially in _preprocess_all() — they do not conflict. The Malayalam normalizer is a no-op for non-Malayalam text, and the general normalizer does not cover Malayalam.

Diff is clean — only 5 files changed, all scoped to the feature:

• omnivoice/models/omnivoice.py — hook into preprocessing
• omnivoice/utils/text_normalization/__init__.py — language dispatcher
• omnivoice/utils/text_normalization/malayalam.py — core normalizer (numbers, currency ₹, %, units)
• tests/test_malayalam_normalizer.py — 327 lines of tests

Ready for review

---