ghananlpcommunity/Ghana_English-Twi_Code-switching_Speech-ipa

KasaSpeech English–Twi Code-Switching Speech — IPA

A phonemised version of ghananlpcommunity/Ghana_English-Twi_Code-switching_Speech (KasaSpeech) with one added column: ipa.

Every other column — audio included — is carried over byte-for-byte, and row order is unchanged, so this dataset aligns one-to-one with the original.

The ipa column

Each transcript is converted to a phoneme sequence with ghanag2p-uni, the Twi-only grapheme-to-phoneme library built on ghana-g2p, run with normalise=True.

One language, applied throughout. ghanag2p-uni phonemises everything under Twi rules — there is no language switch and no per-token language detection. Because KasaSpeech is code-switched, the English stretches are rendered as a Twi speaker's phonology rather than with an English G2P. That is the intended behaviour, and it is what makes the column consistent, but it is the key thing to understand before training on it.

Phoneme units are separated by single spaces, so the column tokenises by splitting on whitespace. Arabic numerals are expanded to English words before conversion ("5" → five). Punctuation is dropped.

transcript	ipa
Ɛyɛ den oo, but I’ll try.	ɛ j ɛ d e n o b u t ɪ ʔ l l t ɾ j
"Nsuo no asa? I’m feeling very thirsty."	n s u o n o a s a ɪ ʔ m f e e l ɪ ŋ v e ɾ j t h ɪ ɾ s t j
"Laptop no ayɛ hye dodo, turn it off."	l a pʰ t o pʰ n o a j ɛ ʃ e d o d o t u ɾ n ɪ t o f f

Two rows in the source whose transcript is punctuation only (]) have an empty ipa string.

Aligned to an existing phoneme inventory

The phonemes are not raw Twi IPA. They are mapped onto the 79-token inventory of KoelLabs/xlsr-english-01, so this column can be used as a CTC target for that model with no vocabulary extension and no resized head. Across all 2,676,779 phoneme tokens here, 100% fall inside that inventory — 33 of its tokens are used, and <unk> cannot occur.

Getting there is lossy, by design:

t͡ʃ d͡ʒ	tie bar dropped — same phoneme, written untied
tʰ → t	the inventory has kʰ and pʰ, but no tʰ
oː iː	vowel length is not contrastive there
ç → ʃ, ɲ → n j	Twi hy and ny have no counterpart
kʷ → k w, hʷ → h w	labialisation becomes a glide
ie → i e, oɔ → o ɔ	vowel clusters split into their parts

If you want faithful Twi IPA rather than model-aligned phonemes, regenerate the column with GhanaG2PUni(normalise=False) from the transcript field.

Usage

from datasets import load_dataset, Audio

ds = load_dataset(
    "ghananlpcommunity/Ghana_English-Twi_Code-switching_Speech-ipa",
    split="train",
)
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

sample = ds[0]
print(sample["transcript"])
print(sample["ipa"].split())   # phoneme tokens

Reproduce the column yourself:

from ghanag2p_uni import GhanaG2PUni

g = GhanaG2PUni(normalise=True)
g.ipa("Ɛyɛ den oo, but I'll try.", sep=" ")
# 'ɛ j ɛ d e n o b u t ɪ ʔ l l t ɾ j'

Splits

Split	Samples	Duration
Train	50,965	83.94 hours
Validation	2,159	6.80 hours
Test	1,731	4.84 hours
Total	54,855	95.58 hours

Fields

Field	Type	Description
speaker_id	string	Anonymous speaker identifier
age_range	string	Speaker age group
gender	string	Speaker gender
prompt_set	string	Prompt category used during recording
transcript	string	Human-annotated English–Twi code-switched transcript
duration	float32	Audio duration in seconds
split	string	Dataset split label
audio	Audio	Speech recording (stored at 48 kHz)
file_name	string	Original audio filename
error	string	Optional annotation or recording error label
ipa	string	Space-separated phonemes of transcript, in the KoelLabs inventory

Notes carried over from the source

• Audio is stored at 48 kHz; cast to 16 kHz for most speech models.
• The split column does not always agree with the file a row lives in — trust the split you loaded, not the column.
• Demographic representation is not perfectly balanced, and recording conditions vary across devices and environments.
• Minor transcription inconsistencies may remain.

Credit

All speech, transcripts and metadata are the work of the original KasaSpeech authors; this derivative adds only the ipa column.

@dataset{kasaspeech2026,
  title={KasaSpeech: A Large-Scale English--Twi Code-Switching Speech Dataset},
  author={Dotse, Kenneth},
  year={2026},
  url={https://huggingface.co/datasets/Kennethdot/Ghana_English-Twi_Code_switching_ASR}
}

Licence: Apache-2.0, as the source dataset.