buchwandler

Arabic MSA

KokoroG2P provides a native Arabic frontend for Modern Standard Arabic (MSA). It is not a dialect frontend and does not identify or convert Arabic dialects.

Construction

from kokorog2p import get_g2p, phonemize

g2p = get_g2p("ar", diacritizer="none")
result = phonemize(
    "مَرْحَبًا بِكَ؟",
    language="ar",
    g2p=g2p,
    return_phonemes=True,
    return_ids=True,
)

The native frontend uses the existing Arabic eSpeak-ng voice and raw IPA. It then applies a small Nabra cleanup policy and targets nabra-82m-v0.1. The Nabra profile reserves IDs 7 and 8 for ʕ and ħ; those IDs are intended for the Nabra model, not stock Kokoro 1.0.

Diacritization

Already-vocalized text can use diacritizer="none" and does not require CAMeL Tools. Unvocalized text can use diacritizer="auto" (the default), which uses CAMeL when its local package and data are available and otherwise emits a warning and continues. Use diacritizer="camel-tools" with strict_diacritizer=True when missing dependencies or data must fail.

CAMeL and its MSA data are optional. KokoroG2P never downloads or provisions that data. Install the optional package and provision disambig-mle-calima-msa-r13 using CAMeL Tools’ documented workflow.

Adapters can be injected when a different local diacritizer is required. Each adapter must return one output token for every input Arabic lexical token.

Source behavior and offsets

  • Arabic harakat remain attached to their lexical source token.

  • ASCII Latin runs are suppressed by default with latin_policy="drop".

  • Numeric square-bracket citation spans such as [12] are suppressed.

  • Real parentheses remain distinct and are not treated as citations.

  • Source token offsets always refer to the original input string.

  • ،, ؛, and ؟ are mapped to ,, ;, and ? after source classification.

Arabic numbers, dates, currencies, and dialect-specific normalization are not implemented by this frontend. Those belong to the upstream semantic normalization architecture.

Generic eSpeak backend

get_g2p("ar", backend="espeak") intentionally returns the generic EspeakOnlyG2P. It does not select the Nabra cleanup or vocabulary profile.