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.

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

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:

git clone https://github.com/GhanaNLP/ghana-g2p
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. The naw rules contain no p; bud has no e or o; ada has no h or r. Ewe's rules map orthographic <y> to IPA y — a front rounded vowel — where it is the palatal glide /j/, as it is in every other language in the set. Missing letters are filled from a patch table of conventional Ghanaian readings; the Ewe error is corrected outright.

Provenance is part of the output

A donor language gives plausible phonemes, not authoritative ones. Every result says where its rules came from, so you can filter on it rather than guess:

r = GhanaG2P("Sehwi").convert("kɔ ekyi")
r.phonemes    # the phonemes
r.tier        # 'donor'
r.rules       # 'any'  (Anyin — closely related Bia language)
r.is_donor    # True
r.dropped     # characters no rule could resolve

Tiers are:

GhanaG2P(...).info carries the family, measured coverage, and a note explaining each non-native choice.

Languages

42 languages. Names, ISO codes, common alternates and ghana-speech config names all resolve to the same entry, so you don't have to know the ISO code:

GhanaG2P("Asante Twi")        # by name
GhanaG2P("twi")               # by code
GhanaG2P("Akuapem_Twi_twi")   # by ghana-speech config name
GhanaG2P("Frafra")            # by common alternate  -> gur (Ninkare)

ghana-g2p --list              # every language, donors marked
ghana-g2p --info Sehwi        # provenance for one language

Donor mappings

Nawuri, Dangme and Bassar also use donors — not for lack of rules, but because their own rule sets are missing core letters (p, h/r, and e/o respectively). They map to Nkonya, Ga and Tem, each within the correct family.

Output format

ipa() and grapheme() strip punctuation and whitespace. By default units run together; pass sep=" " to keep the boundaries.

Use sep=" " unless you have a reason not to. Many units are more than one character — ny kp gb nw kʰ k͡p — so the run-together form is ambiguous about where one phoneme ends and the next begins. Twi nw is a single labialised nasal; a consumer that splits on characters reads it as n + w.

This matters most for forced alignment. With ctc-forced-aligner, feed space-separated units and align with --split_size word, so each phoneme unit gets one timestamped span:

from ghana_g2p import GhanaG2P
GhanaG2P("Asante Twi").ipa("Onyankopɔn nwoma", sep=" ")
# 'o ɲ a n kʰ o pʰ ɔ n nʷ o m a'   -> 13 units, 13 alignment spans

Run-together text would force --split_size char, which splits every multi-character unit. (The aligner's normaliser folds modifier letters — kʰ→kh, nʷ→nw — but unit boundaries survive because they are space-separated.)

Keeping punctuation

Punctuation is stripped by default. Pass punctuation=True to keep it, with each mark as its own unit — attaching it to a neighbouring phoneme would reintroduce the boundary ambiguity that separating units avoids:

GhanaG2P("twi").ipa("Wo ho te sɛn?", sep=" ", punctuation=True)
# 'w o h o tʰ e s ɛ n ?'

Useful when the marks carry prosody you want to model. Bracket pairs left empty once their contents were dropped are removed, so the verse-number convention (23) does not leave a bare ( ); brackets around real content are kept.

Two more notes:

• Apostrophes are not punctuation in several of these orthographies — in Anyin and the Guang languages they mark glottal stop, and are phonemised as ʔ.
• Digits are dropped rather than verbalised; normalise numbers to words before phonemising if you need them spoken.

Batch use

batch() caches per word, which matters on speech corpora where transcriptions repeat:

GhanaG2P("Dagbani").batch(list_of_transcriptions)

Relationship to africa-g2p

This is a wrapper, not a fork. africa-g2p does the segmentation and holds the rule data; ghana-g2p adds the Ghanaian registry, the donor mappings, the normalisation and patch layers, and the provenance reporting. Fixes that belong upstream are reported there.

Licence

Apache-2.0. Underlying rule data is subject to africa-g2p's own licence.

Upstream status

The defects found while building this were reported as AfriSpeech/africa-g2p#2 and fixed upstream in #3. Because the tests assert correct output rather than the presence of a workaround, they kept passing across the change.

Now redundant (fixed upstream, kept only so older africa-g2p checkouts still work): phonetic brackets, the Ga Greek epsilon, and the Ewe <y> mapping.

Still needed here: the Ninkare vowel-length correction, which is a per-language linguistic judgement rather than a data defect, and the missing-letter patch table.

One upstream fix improved output beyond what this wrapper could do on its own — rule keys containing combining marks are now reachable, so Ninkare ẽ yields ɛ̃ (nasalization lowers the vowel) instead of e plus a tilde.