GhanaNLP/ghana-g2p

Grapheme-to-phoneme conversion for Ghanaian languages: 42 languages, IPA and native-orthography output, built on africa-g2p # ghana-g2p Grapheme-to-phoneme conversion for **Ghanaian languages**, built on africa-g2p. africa-g2p covers 400+ African languages from a single general registry. `ghana-g2p` narrows that to the languages actually spoken in Ghana and fixes what breaks when you run real Ghanaian text through it: missing rule sets, Unicode codepoint variants, and a handful of incorrect or incomplete mappings. ```python from ghana_g2p import GhanaG2P tw = GhanaG2P("Asante Twi") tw.ipa("Mfiase no Onyankopɔn bɔɔ ɔsoro.") # 'mfiasenooɲankʰopʰɔnbɔɔɔsoɾo' tw.grapheme("Mfiase no Onyankopɔn bɔɔ ɔsoro.") # 'mfiasenoonyankopɔnbɔɔɔsoro' tw.ipa("Akwaaba", sep=" ") # 'a kʷ a a b a' ``` ## Install ```bash pip install ghana-g2p ``` That pulls africa-g2p, whose rule tables this builds on. Nothing else is needed — both are pure Python with no other runtime dependencies. For development: ```bash git clone github.com cd ghana-g2p pip install -e . ``` ## Why this exists Running the full ghana-speech corpus through africa-g2p surfaced three classes of problem. Each is fixed here, and every fix is recorded as data you can inspect rather than hidden in code. **1. Ten Ghanaian languages have no rules at all.** Birifor, Buli, Konni, Lelemi, Ntrubo, Sehwi, Sekpele, Selee, Tampulma and Tuwuli are simply absent. Each is mapped to a related language chosen by linguistic classification first, then measured orthographic coverage — never by character overlap alone, which ranks a Mande language above Dagaare for Birifor purely because they share the Latin alphabet. **2. Codepoint variants.** Ghanaian orthographies write the same vowel with different Unicode characters. Birifor and Tem spell /ʊ/ as `ʊ` (U+028A); the matching rule sets use `ʋ` (U+028B). Tumulung Sisaala writes `ɩ ʋ` where the `ssl` rules expect `ɪ ʊ`. Normalising these lifts Tem from 0.97 to full coverage and Sisaala from 0.80 to 1.00. **3. Incomplete and incorrect rule sets.** …