---
license: cc-by-nc-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
language:
- tw
tags:
- health
- medical
- speech
- ghana
- african-languages
- phonemes
- ipa
- g2p
pretty_name: Twi Health Speech Dataset Gemini (500 hours) — IPA
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Twi Health Speech Dataset Gemini (500 hours) — IPA

A phonemised version of
[ghananlpcommunity/twi-health-asr-gemini-500hrs](https://huggingface.co/datasets/ghananlpcommunity/twi-health-asr-gemini-500hrs)
with one added column: **`ipa`**.

`audio`, `transcription` and row order are carried over unchanged, so this
dataset aligns one-to-one with the original. ~59k utterances, ~500 hours.

## The `ipa` column

Each `transcription` is converted to a phoneme sequence with
[ghanag2p-uni](https://github.com/michsethowusu/ghanag2p-uni), the Twi-only
grapheme-to-phoneme library built on
[ghana-g2p](https://github.com/GhanaNLP/ghana-g2p), run with `normalise=True`.

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.

### 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](https://huggingface.co/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. Every unit the Twi rules can emit is covered, so
`<unk>` cannot occur.

The mapping 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 instead, regenerate the column from
`transcription` with `GhanaG2PUni(normalise=False)`.

**One language, applied throughout.** `ghanag2p-uni` phonemises everything under
Twi rules — there is no language switch. Where the source transcripts carry
English words, they are rendered as a Twi speaker's phonology rather than with
an English G2P.

## Usage

```python
from datasets import load_dataset, Audio

ds = load_dataset(
    "ghananlpcommunity/twi-health-asr-gemini-500hrs-ipa",
    split="train",
)
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

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

Reproduce the column yourself:

```python
from ghanag2p_uni import GhanaG2PUni

g = GhanaG2PUni(normalise=True)
g.ipa("Mfiase no Onyankopɔn bɔɔ ɔsoro.", sep=" ")
```

## Fields

| Field | Type | Description |
|---|---|---|
| `audio` | `Audio` | Speech recording, 16 kHz |
| `transcription` | `string` | Twi transcript |
| **`ipa`** | `string` | **Space-separated phonemes of `transcription`, in the KoelLabs inventory** |

## Credit

All audio and transcripts are the work of the original dataset's authors; this
derivative adds only the `ipa` column.

Original: https://huggingface.co/datasets/ghananlpcommunity/twi-health-asr-gemini-500hrs

Licence: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), as the
source dataset.