File size: 2,851 Bytes
aca68d3
b531acf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aca68d3
b531acf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
---
language:
- multilingual
language_bcp47:
- akp-AK
- any-Iv
- twi-GH
- avn-GH
- bqa-GH
- bim-GH
- biv-GH
- bib-GH
- bwu-GH
- ncu-GH
- dga-GH
- dag-GH
- adq-GH
- mzw-GH
- ewe-GH
- fat-GH
- ffm-GH
- acd-GH
- gjn-GH
- hau-GH
- kbp-GH
- xsm-GH
- xon-GH
- kma-GH
- kus-GH
- lef-GH
- maw-GH
- naw-GH
- Gur-GH
- nko-GH
- ntr-GH
- nzi-GH
- sig-GH
- sfw-GH
- lip-GH
- snw-GH
- sil-GH
- siwu-GH
- tpm-GH
- kdh-GH
- bov-GH
- vag-GH
tags:
- speech
- multilingual
- african-languages
- ghana
- phonemes
- tts
---

# Ghana Speech (Phonemized)

Preprocessed multilingual speech dataset from
[`ghananlpcommunity/ghana-speech`](https://huggingface.co/datasets/ghananlpcommunity/ghana-speech),
covering **42 Ghanaian and West African languages** with IPA phoneme transcriptions.

This dataset contains only the **metadata and phonemes** — no audio.
Use it together with the source dataset to load audio on-the-fly.

## Contents

Each language is a separate **config** (Parquet). Columns:

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique clip ID (matches audio IDs in the source dataset) |
| `text` | string | Raw orthographic transcription |
| `phon` | string | IPA phoneme transcription (espeak `lfn` voice) |
| `lang` | string | Language subset name (e.g. `Akuapem_Twi_twi`) |
| `lang_id` | int | Integer language ID (0-41) for speaker-conditioning slot |
| `dur` | float | Clip duration in seconds |
| `source_file` | string | Original source filename |
| `split` | string | `train` or `val` |

Repo-level files:

- **`lang_map.json`** — language name to lang_id mapping
- **`ghana_speech.json`** — mel normalisation stats (mel_mean, mel_std)
- **`lang_stats.json`** — per-language clip counts
- **`val_ids.json`** — list of validation clip IDs

## Filtering criteria

- Duration: 1.0s to 15.0s
- Phonemes: at least 2 non-space characters after phonemization
- Empty text: excluded

## Validation split

32 clips per language, stratified. Val clips have `split="val"` in the Parquet configs.

## Usage with Matcha-TTS

```python
from datasets import load_dataset, Audio

# Load phonemes for a specific language
ds = load_dataset("ghananlpcommunity/ghana-speech-phonemized", "Akuapem_Twi_twi", split="train")

# Load audio from the source dataset on-the-fly
audio_ds = load_dataset("ghananlpcommunity/ghana-speech", "Akuapem_Twi_twi", split="train")

# Combine: use phonemes from this dataset, audio from source
for phon_row, audio_row in zip(ds, audio_ds):
    phonemes = phon_row["phon"]
    audio = audio_row["audio"]["array"]
```

## Mel normalisation stats

```json
{"mel_mean": -5.728251241639019, "mel_std": 3.3875457364162376}
```

## Statistics

- **Total clips**: ~1,195,000
- **Languages**: 42
- **Train clips**: ~1,194,000
- **Val clips**: 1,344 (32 per language)

## License

Same as source dataset.