dhvani 0.2.0

Phonetic normalization for Hinglish text. Resolves spelling variation in Romanized Hindi using IPA as a bridge representation.

dhwani

Phonetic normalization for Hinglish text.

dhwani (ध्वनि = "sound") understands that "bahut", "bohot", "boht", and "bhot" are all the same word. It normalizes the chaos of Romanized Hindi into something computers can actually work with.

Why?

600M+ Indians write online in Hinglish (Hindi in Latin script mixed with English). But there's no standard spelling:

"बहुत" gets written as: bahut, bohot, boht, bhot, bahot
"अच्छा" gets written as: accha, achha, acha, achaa
"कैसे" gets written as: kaise, kese, kayse

Every NLP tool breaks on this. dhwani fixes it.

Install

pip install git+https://github.com/Kkoundinyaa/dhwani.git

For higher accuracy on rare words (optional):

pip install "dhwani[models] @ git+https://github.com/Kkoundinyaa/dhwani.git"

Usage

import dhwani

# Check if two words are the same (variant spellings)
dhwani.are_same("bahut", "bohot")   # True
dhwani.are_same("accha", "achha")   # True
dhwani.are_same("bahut", "accha")   # False

# Convert Hinglish to Devanagari
dhwani.to_devanagari("bohot accha movie thi yaar")
# -> "बहुत अच्छा movie थी यार"

# Convert to IPA (phonetic representation)
dhwani.to_ipa("kaise ho bhai")
# -> "kɛːseː ɦoː bʱaːiː"

# Word-level language identification
dhwani.identify_languages("ye movie really acchi thi bro")
# -> [("ye", "hi"), ("movie", "en"), ("really", "en"), ("acchi", "hi"), ("thi", "hi"), ("bro", "hi")]

# Normalize text
dhwani.normalize("bohot acha movie thi")
# -> canonical normalized form

CLI

dhwani devanagari "bohot accha movie thi yaar"
# बहुत अच्छा movie थी यार

dhwani ipa "kaise ho bhai"
# kɛːseː ɦoː bʱaːiː

dhwani same "bahut" "bohot"
# True (phonetic similarity: 1.00)

dhwani langs "ye movie bohot acchi thi"
# ye[hi] movie[en] bohot[hi] acchi[hi] thi[hi]

How It Works

dhwani routes through IPA (International Phonetic Alphabet) as a bridge representation. All variant spellings of a word produce the same sound, so they map to the same IPA:

"bahut" ─┐
"bohot" ─┤──> /bəɦʊt̪/ ──> बहुत
"boht"  ─┤
"bhot"  ─┘

Three-tier architecture for speed:

Tier	Method	Speed	Coverage
1	Lexicon lookup (151K entries)	0.001ms	~95% of common words
2	AI model (IndicXlit + epitran)	~4s	Handles anything
3	Rule-based G2P	0.005ms	Always available

Plus a runtime cache that learns: words processed by Tier 2 get cached permanently, so the library gets faster over time.

Features

• Phonetic equivalence: Detect if two words are the same regardless of spelling
• Transliteration: Romanized Hindi to Devanagari (and back)
• IPA conversion: Any Hindi text (Roman or Devanagari) to IPA
• Language ID: Word-level Hindi/English classification in mixed text
• Zero dependencies for basic use (lexicon + rules)
• 151K-word lexicon built from real Hindi corpora
• Runtime learning: Gets smarter the more you use it

Research

Built on findings from IPA-GPT research at Ohio State University, which showed that phonetic (IPA) representations dramatically improve cross-lingual NLP for script-divergent languages like Hindi-Urdu.

License

MIT