ghana-speech-phonemes

IPA and native-orthography phoneme transcriptions for every utterance in ghananlpcommunity/ghana-speech.

This is a text-only companion: it carries the transcriptions and their phonemisations, not the audio. Join it to the source dataset on id to get audio alongside phonemes. Rows and ids correspond one-to-one with the source.

Phonemes were produced with ghana-g2p, which wraps africa-g2p with a Ghanaian language registry and corrections for the orthographic and rule-data problems that surfaced while processing this corpus.

Columns

column	description
id	utterance id, matching the source dataset
language	language name
text	original transcription
ipa_phonemes_spaced	IPA phoneme units, space-separated — use this one
grapheme_phonemes_spaced	native-orthography units, space-separated
ipa_phonemes_spaced_punct	same as above, with punctuation kept as its own units
grapheme_phonemes_spaced_punct	same, native orthography
ipa_phonemes	the same IPA units run together, no separators
grapheme_phonemes	the same native units run together
phoneme_count	number of phoneme units in the utterance
duration	audio length in seconds
source_file	source file reference
phoneme_rules	the africa-g2p rule set actually used
phoneme_tier	native, equivalent, or donor — see below

Phoneme format

Phonemes come in three shapes: space-separated units, the same run together, and space-separated units with punctuation retained.

text                 Na Misraimni no kura ntomanwomfoɔ nsadua.
ipa_phonemes_spaced  n a m ɪ s ɾ a ɪ m n ɪ n o kʰ u ɾ a n tʰ o m a nʷ o m f o ɔ n s a d u a
ipa_phonemes         namɪsɾaɪmnɪnokʰuɾantʰomanʷomfoɔnsadua

Prefer the _spaced columns. Many units are more than one character — ny, kp, gb, nw, kʰ, k͡p — so the run-together form does not record where one phoneme ends and the next begins. Twi nw is a single labialised nasal nʷ; a consumer that splits on characters reads it as two sounds. The collapsed columns are exactly the spaced ones with separators removed, kept for uses that genuinely want a bare symbol string.

Forced alignment

This dataset is built for phoneme-level alignment with ctc-forced-aligner. Feed the spaced column and align at word granularity, so each phoneme unit gets one timestamped span:

ctc-forced-aligner --audio_path utt.wav --text_path phonemes.txt \
                   --language twi --split_size word

--split_size word splits on whitespace, so one "word" is one phoneme unit. Using the collapsed column would force --split_size char, which breaks every multi-character unit apart. The aligner's normaliser folds modifier letters on the way in (kʰ→kh, nʷ→nw), but unit boundaries survive because they are space-separated — phoneme_count tells you how many spans to expect.

Punctuation

The _punct columns keep punctuation, each mark as its own space-separated unit:

ipa_phonemes_spaced        w o h o tʰ e s ɛ n
ipa_phonemes_spaced_punct  w o h o tʰ e s ɛ n ?

Use them where the marks carry prosody — pauses and phrase breaks — for instance training a TTS front-end. For forced alignment prefer the plain _spaced columns: a punctuation mark has no acoustic realisation, so giving it its own span asks the aligner to place something that was never spoken. (The aligner's own normaliser strips punctuation anyway.)

Marks are separate units rather than attached to a neighbouring phoneme, so one token is still one unit. Bracket pairs left empty once digits were dropped are removed, so the verse number (23) does not leave a bare ( ); brackets around real content are kept.

Two more things to know about the format:

• Apostrophes are not punctuation in several of these orthographies. In Anyin and the Guang languages they mark glottal stop, and Dagbani writes elision the same way (Filistianim’). They are phonemised as ʔ and appear in every column, including the ones with punctuation stripped. In the grapheme_* columns the unit is written ', which looks like punctuation but is a phoneme — so do not filter those columns by Unicode punctuation category. A quotation mark detached from a word is punctuation and is handled as such.
• Digits are dropped, not verbalised. 10,748 rows (0.76%) consist only of digits or punctuation — mostly verse numbers — and have empty phoneme strings. Filter them out, or normalise numbers to words before phonemising, if that matters for your use.

Read the phoneme_tier column

africa-g2p does not have rules for every Ghanaian language, and some of the rule sets it does have are incomplete. Rather than hide that, every row records where its phonemes came from:

tier	meaning	rows
native	the language's own rules	903,908
equivalent	the same language under a different code (Dagaare is dgd, Deg is mfi)	220,494
donor	a related language's rules, because none exist for this one	287,065

donor phonemes are plausible, not authoritative. Sehwi through Anyin rules is close — they are sister Bia languages — but you should not treat donor-derived phonemes as a substitute for a phonemic description of the language. Filter on phoneme_tier == 'native' if you need only first-hand rules.

Donor choices were made by linguistic classification first and measured orthographic coverage second. Coverage alone is a poor selector: for Birifor it ranks an unrelated Mande language above Dagaare purely because they share the Latin alphabet.

Per-language provenance

config	language	rules	tier	coverage	rows
Akuapem_Twi_twi	Twi	twi	native	1.0000	52,650
Anyin_any	Anyin	any	native	1.0000	7,968
Asante_Twi_twi	Twi	twi	native	1.0000	143,383
Avatime_avn	Avatime	avn	native	1.0000	9,956
Bassar_Ntcham_bud	Bassar	kdh	donor	1.0000	56,096
Bimoba_bim	Bimoba	bim	native	1.0000	46,911
Birifor_Southern_biv	Southern Birifor	dgd	donor	1.0000	7,942
Bissa_bib	Bissa	bib	native	1.0000	8,799
Buli_bwu	Buli	dgd	donor	1.0000	39,605
Chumburung_ncu	Chumburung	ncu	native	0.9988	9,332
Dagaare_dga	Dagaare	dgd	equivalent	1.0000	7,369
Dagbani_dag	Dagbani	dag	native	1.0000	46,526
Dangme_ada	Dangme	gaa	donor	1.0000	8,283
Deg_mzw	Deg	mfi	equivalent	0.9999	73,947
Ewe_ewe	Ewe	ewe	native	1.0000	108,240
Fante_fat	Fante	fat	native	1.0000	60,933
Fulfulde_Maasina_ffm	Maasina Fulfulde	ffm	native	1.0000	11,895
Gikyode_acd	Gikyode	acd	native	1.0000	9,377
Gonja_gjn	Gonja	gjn	native	1.0000	53,590
Hausa_hau	Hausa	hau-nigeria	equivalent	1.0000	92,082
Kabiye_kbp	Kabiye	kbp	native	0.9999	12,302
Kasem_xsm	Kasem	xsm	native	1.0000	74,077
Konkomba_xon	Konkomba	xon	native	1.0000	50,337
Konni_kma	Konni	dgd	donor	1.0000	8,753
Kusaal_kus	Kusaal	kus	native	1.0000	48,590
Lelemi_lef	Lelemi	avn	donor	1.0000	59,914
Mampruli_maw	Mampruli	maw	native	1.0000	8,961
Nawuri_naw	Nawuri	nko	donor	1.0000	10,495
Ninkare_gur	Ninkare	gur	native	1.0000	17,149
Nkonya_nko	Nkonya	nko	native	0.9998	13,878
Ntrubo_ntr	Ntrubo	kbp	donor	1.0000	11,660
Nzema_nzi	Nzema	nzi	native	1.0000	9,617
Paasaal_sig	Paasaal	sig	native	1.0000	8,384
Sehwi_sfw	Sehwi	any	donor	1.0000	46,878
Sekpele_lip	Sekpele	ewe	donor	1.0000	8,834
Selee_snw	Selee	akp	donor	1.0000	8,859
Sisaala_Tumulung_sil	Tumulung Sisaala	sld	equivalent	1.0000	47,096
Siwu_akp	Siwu	akp	native	1.0000	11,837
Tampulma_tpm	Tampulma	xsm	donor	1.0000	8,908
Tem_kdh	Tem	kdh	native	1.0000	11,482
Tuwuli_bov	Tuwuli	avn	donor	1.0000	10,838
Vagla_vag	Vagla	vag	native	1.0000	67,734

Coverage is the share of characters in a sample of real transcriptions that the rule set could resolve. The residual below 1.0 is combining tone diacritics, not segments.

Known corrections applied

Processing this corpus surfaced defects in the upstream rule data. All were reported and fixed upstream (africa-g2p#3, #4):

• Ewe mapped orthographic <y> to IPA y (a front rounded vowel) rather than the palatal glide /j/.
• Ninkare and Kabiye rule values carried phonetic brackets — every value was written [a], [k͡p] and so on.
• Ga emitted Greek small letter epsilon (U+03B5) where IPA open-e is U+025B.
• Ninkare marked e, i and o long while a, u, ɛ, ɔ stayed short — the source chart used length to render the ATR contrast.
• Codepoint variants: Birifor and Tem spell /ʊ/ U+028A where the matching rule sets use U+028B; Tumulung Sisaala writes ɩ ʋ where the ssl rules expect ɪ ʊ. Normalising these lifted Sisaala from 0.80 to 1.00 coverage and Tem from 0.97 to 1.00.
• Missing letters: the naw rules contain no p, bud no e or o, ada no h or r. Those languages use rules from a same-family relative instead.
• Unreachable rule entries: rule keys containing a combining mark could never match, so the engine fell back to the base letter — 1,395 entries across 193 languages. Chumburung writes the ATR contrast with a combining macron below (e̱ o̱), and that distinction was being dropped entirely. Fixed upstream; 99.9% of Chumburung rows changed as a result.

Usage

from datasets import load_dataset

ds = load_dataset("ghananlpcommunity/ghana-speech-phonemes", "Asante_Twi_twi", split="train")
ds[0]["ipa_phonemes_spaced"]        # 'o ɲ a n kʰ o pʰ ɔ n'  <- prefer this
ds[0]["ipa_phonemes_spaced_punct"]  # same, punctuation kept
ds[0]["ipa_phonemes"]               # 'oɲankʰopʰɔn'  (unit boundaries not recorded)

Joining to the audio:

from datasets import load_dataset

audio = load_dataset("ghananlpcommunity/ghana-speech", "Asante_Twi_twi", split="train")
phon  = load_dataset("ghananlpcommunity/ghana-speech-phonemes", "Asante_Twi_twi", split="train")
lookup = dict(zip(phon["id"], phon["ipa_phonemes"]))
audio = audio.map(lambda r: {"ipa_phonemes": lookup.get(r["id"], "")})

Phonemising your own text:

pip install git+https://github.com/GhanaNLP/ghana-g2p

from ghana_g2p import GhanaG2P
GhanaG2P("Asante Twi").ipa("Akwaaba")

Licence

CC BY-NC 4.0, following the source dataset.

Citation

Please cite the source dataset ghananlpcommunity/ghana-speech alongside this one.