Aybee5/hausa-tts-small

Mimic Studio Hausa TTS Dataset

This dataset was created using Mimic Studio for training text-to-speech models.

Dataset Details

• Total Samples: 62
• Language: Hausa
• Speakers: 2 speakers (Surajo Nuhu Umar, Umar Musa Halliru)
• Format: Compatible with Unsloth TTS models
• Audio Format: WAV files, 24kHz sampling rate

Speaker Distribution

Surajo Nuhu Umar     32
Umar Musa Halliru    30

Dataset Structure

Aybee5/hausa-tts-small/
├── data/
│   ├── train.parquet           # Metadata (source, text, audio paths)
│   └── audio_files/
│       ├── 97f373e8-f6e6-.../  # Speaker 1 audio files
│       ├── b0db0a87-2206-.../  # Speaker 2 audio files
│       └── c3621689-ca53-.../  # Additional audio files
└── README.md

The dataset has the following columns:

• source: Speaker name (for multi-speaker training)
• text: Transcription/prompt in Hausa
• audio: Audio file (automatically loaded from data/audio_files/)

Usage

This dataset is designed for use with Unsloth TTS models.

from datasets import load_dataset, Audio

# Load the dataset - audio files are automatically downloaded
ds = load_dataset("Aybee5/hausa-tts-small", split="train")

# Audio is already configured with 24kHz sampling rate
print(ds[0])
# {'source': 'Umar Musa Halliru', 'text': 'Tsarki', 'audio': {'array': [...], 'sampling_rate': 24000, 'path': '...'}}

# Access audio array directly
audio_array = ds[0]['audio']['array']
sampling_rate = ds[0]['audio']['sampling_rate']

For Unsloth TTS Training

from datasets import load_dataset, Audio
from transformers import AutoProcessor
import torch

processor = AutoProcessor.from_pretrained("unsloth/csm-1b")
raw_ds = load_dataset("Aybee5/hausa-tts-small", split="train")

# Audio is already at 24kHz, no need to recast
speaker_key = "source"

def preprocess_example(example):
    conversation = [
        {
            "role": str(example[speaker_key]),
            "content": [
                {"type": "text", "text": example["text"]},
                {"type": "audio", "path": example["audio"]["array"]},
            ],
        }
    ]

    model_inputs = processor.apply_chat_template(
        conversation,
        tokenize=True,
        return_dict=True,
        output_labels=True,
        text_kwargs={
            "padding": "max_length",
            "max_length": 256,
            "pad_to_multiple_of": 8,
            "padding_side": "right",
        },
        audio_kwargs={
            "sampling_rate": 24_000,
            "max_length": 240001,
            "padding": "max_length",
        },
        common_kwargs={"return_tensors": "pt"},
    )

    required_keys = ["input_ids", "attention_mask", "labels", "input_values", "input_values_cutoffs"]
    processed_example = {key: model_inputs[key][0] for key in required_keys}
    
    return processed_example

# Process the dataset
processed_ds = raw_ds.map(
    preprocess_example,
    remove_columns=raw_ds.column_names,
)

print(f"Processed {len(processed_ds)} samples")

Citation

If you use this dataset, please cite the Mimic Studio project:

@misc{mimicstudio,
  title = {Mimic Recording Studio},
  author = {Mycroft AI},
  howpublished = {\url{https://github.com/MycroftAI/mimic-recording-studio}},
  year = {2019}
}

License

MIT License

Dataset Creation

This dataset was created by:

1. Recording audio using Mimic Studio
2. Extracting data from the SQLite database
3. Converting to relative paths
4. Organizing into HuggingFace standard structure (data/ folder)
5. Configuring proper audio feature loading

The audio files are stored in data/audio_files/ and are automatically downloaded when loading the dataset.