Datasetpublic

Alkamal01/suda-hausa-tts

Suda Hausa TTS Dataset High-quality Hausa speech dataset prepared for fine-tuning XTTS-v2 as part of the Suda project — an open-source Hausa voice AI. Dataset Details Property Value Language Hausa (ha) Clips 4,960 Duration ~10 hours Sample rate 22,050 Hz mono WAV Format LJSpeech (metadata.csv + wavs/) Speaker Single speaker (male, Nigerian Hausa) Source BibleTTS Hausa License CC-BY-SA 4.0 Source Derived from BibleTTS… See the full description on the dataset page: https://huggingface.co/datasets/Alkamal01/suda-hausa-tts.

sourceHugging Facecc-by-sa-4.0updated 5mo agoView on Hugging Face

0likes158downloads

Dataset Card

Suda Hausa TTS Dataset

High-quality Hausa speech dataset prepared for fine-tuning XTTS-v2 as part of the Suda project — an open-source Hausa voice AI.

Dataset Details

Property	Value
Language	Hausa (`ha`)
Clips	4,960
Duration	~10 hours
Sample rate	22,050 Hz mono WAV
Format	LJSpeech (`metadata.csv` + `wavs/`)
Speaker	Single speaker (male, Nigerian Hausa)
Source	BibleTTS Hausa
License	CC-BY-SA 4.0

Source

Derived from BibleTTS Hausa by Meyer et al. (2022):

—Original: 40,603 verses, 86.6 hours, studio-quality 48kHz FLAC, single speaker
—This subset: 4,960 clips (1–30s), resampled from 48kHz → 22,050 Hz mono WAV

Structure

suda-dataset/
├── metadata.csv      # pipe-delimited: wavs/file.wav|transcript
└── wavs/
    ├── suda_0000000.wav
    ├── suda_0000001.wav
    └── ...

metadata.csv format (LJSpeech):

wavs/suda_0000000.wav|Zan miƙa hannuna gāba da Yahuda da kuma dukan mazaunan Urushalima.
wavs/suda_0000001.wav|Aka yi shela a duk fāɗin Yahuda da Urushalima.

Usage

Load metadata

python

import pandas as pd

df = pd.read_csv(
    "hf://datasets/Alkamal01/suda-hausa-tts/metadata.csv",
    sep="|",
    header=None,
    names=["path", "transcript"],
)
print(df.head())

Fine-tune XTTS-v2

python

from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.datasets import load_tts_samples

config = XttsConfig()
config.datasets = [{
    "formatter": "ljspeech",
    "dataset_name": "suda_hausa",
    "path": "/path/to/suda-dataset",
    "meta_file_train": "metadata.csv",
    "language": "ha",
}]

See full training script: train_xtts_hausa.py

Processing

Dataset was prepared using open-source scripts in the OribAI-all repo:

1.Streamed BibleTTS Hausa parquet shards from HuggingFace
2.Filtered clips to 1–30 seconds
3.Resampled 48kHz FLAC → 22,050 Hz mono WAV (scipy.signal.resample_poly)
4.Wrote LJSpeech-format metadata.csv

Citation

If you use this dataset, cite the original BibleTTS paper:

bibtex

@article{meyer2022bibletts,
  title     = {BibleTTS: a large, high-fidelity, multilingual, and uniquely African speech corpus},
  author    = {Meyer, Josh and others},
  journal   = {arXiv preprint arXiv:2207.03546},
  year      = {2022}
}

Project

Part of Suda — open-source Hausa voice AI built by Keuro Lab.