Hindi G2P

Installation

python -m pip install "kokorog2p[hi]"

The extra uses the existing eSpeak integration. No Hindi Lexphon or G2Lex data is bundled or required.

Supported aliases

The aliases hi, hi-in, hi_IN, hin, and hindi normalize to canonical language hi-in. All aliases share the same cached HindiG2P instance when factory options match.

from kokorog2p import get_g2p, phonemize_prepared

g2p = get_g2p("hi")
assert g2p is get_g2p("hindi")

result = phonemize_prepared(
    "नमस्ते दुनिया!",
    language="hi",
    return_ids=True,
)
assert result.phonemes
assert result.token_ids

Raw IPA and model compatibility

HindiG2P uses the eSpeak-NG hi voice and requests raw IPA. It does not call the English-oriented from_espeak() conversion because that conversion would alter Hindi vowel quality, length, nasalization, and consonant distinctions. The Hindi profile only strips eSpeak language markers and tie controls and removes unsupported U+0329 syllabicity. The final output is validated against the Kokoro v1.0 vocabulary.

strict=True reports an unsupported symbol with the source token, raw IPA, and normalized IPA. With strict=False, an invalid lexical token is returned without phonemes instead of emitting an invalid model input.

Lazy initialization and routing

Factory construction is lazy. Creating get_g2p("hi") does not initialize eSpeak-NG; the backend is created on first pronunciation. Explicit backend="espeak" selects the same Hindi-safe frontend.

Hindi has no runtime lexicon and does not provide positive automatic-routing evidence. Its eSpeak pronunciation is realization data only. Use explicit language or language spans when composing Hindi with other prepared languages.