Hausa TTS Dataset (HausaTTSEmbed)

This dataset contains 1,283 Hausa language audio recordings with transcriptions for Text-to-Speech (TTS) model training.

Dataset Details

• Language: Hausa (ha)
• Total Samples: 1,283
• Speakers: 3 unique speakers
• Audio Format: WAV files
• Sample Rate: Original recordings (will be resampled to 24kHz during training)
• Text Length: 4-141 characters (average: 24 characters)

Dataset Structure

Each example contains:

• audio: Audio file in WAV format
• text: Hausa transcription with proper diacritics (e.g., "Ansamu ɓaraka acikin shirin")
• speaker_id: UUID of the speaker (3 unique values)

Data Fields

{
    'audio': {
        'path': str,      # Path to audio file
        'array': ndarray, # Audio waveform
        'sampling_rate': int
    },
    'text': str,          # Hausa transcription
    'speaker_id': str     # Speaker identifier
}

Usage

Recommended: Download All Files First

To ensure all audio files are available and avoid rate limits, authenticate first:

from huggingface_hub import snapshot_download, HfApi
from datasets import load_dataset, Audio
import os
import time

# IMPORTANT: Login FIRST and WAIT for confirmation
# Method 1: Using token directly (RECOMMENDED for Colab)
from huggingface_hub import login
HF_TOKEN = "hf_YourTokenHere"  # Get from https://huggingface.co/settings/tokens
login(token=HF_TOKEN)

# Verify login worked
api = HfApi()
user_info = api.whoami(token=HF_TOKEN)
print(f"✓ Logged in as: {user_info['name']}")

# Small delay to ensure auth propagates
time.sleep(2)

# Download entire dataset (parquet + all audio files)
print("\nDownloading dataset (~2GB)...")
local_dir = snapshot_download(
    "Aybee5/HausaTTSEmbed",
    repo_type="dataset",
    local_dir="hausa_tts_data",
    token=HF_TOKEN,  # Pass token explicitly
    max_workers=1,   # Reduce concurrent requests to avoid rate limits
    resume_download=True  # Resume if interrupted
)

# Load from downloaded files
dataset = load_dataset(
    "parquet",
    data_files=f"{local_dir}/data/*.parquet",
    split="train"
)

# Fix audio paths to absolute paths
dataset = dataset.map(lambda x: {"audio": os.path.join(local_dir, x["audio"]), **x})

# Cast to Audio type
dataset = dataset.cast_column("audio", Audio(sampling_rate=22050))

print(f"✓ Loaded {len(dataset)} samples")

# Access sample
sample = dataset[0]
print(f"Text: {sample['text']}")
print(f"Audio shape: {sample['audio']['array'].shape}")

Alternative: Interactive Login (prompts for token)

from huggingface_hub import login, snapshot_download
import time

# This will prompt you to paste your token
login()
time.sleep(2)  # Wait for auth to propagate

# Then download
local_dir = snapshot_download(
    "Aybee5/HausaTTSEmbed",
    repo_type="dataset",
    local_dir="hausa_tts_data",
    max_workers=1  # Reduce concurrent requests
)

For Unsloth TTS Training (Complete Code)

Use this complete code in your Unsloth/Colab notebook:

from huggingface_hub import snapshot_download, login, HfApi
from datasets import load_dataset, Audio
import os
import time

# ==================== STEP 1: AUTHENTICATE ====================
# Replace with your actual token from https://huggingface.co/settings/tokens
HF_TOKEN = "hf_YourTokenHere"

print("Authenticating with HuggingFace...")
login(token=HF_TOKEN)

# Verify authentication
api = HfApi()
user_info = api.whoami(token=HF_TOKEN)
print(f"✓ Logged in as: {user_info['name']}\n")

# Wait for auth to propagate
time.sleep(2)

# ==================== STEP 2: DOWNLOAD DATASET ====================
print("Downloading Hausa TTS dataset (~2GB)...")
print("Using reduced concurrency to avoid rate limits...\n")

local_dir = snapshot_download(
    "Aybee5/HausaTTSEmbed",
    repo_type="dataset",
    local_dir="/content/hausa_tts",  # Use /content/ for Colab
    token=HF_TOKEN,  # Pass token explicitly
    max_workers=1,   # Single threaded to avoid rate limits
    resume_download=True
)

print(f"✓ Downloaded to: {local_dir}\n")

# ==================== STEP 3: LOAD DATASET ====================
raw_ds = load_dataset(
    "parquet",
    data_files=f"{local_dir}/data/*.parquet",
    split="train"
)

# ==================== STEP 4: FIX AUDIO PATHS ====================
raw_ds = raw_ds.map(lambda x: {"audio": os.path.join(local_dir, x["audio"]), **x})

# ==================== STEP 5: HANDLE SPEAKERS ====================
speaker_key = "source"
if "source" not in raw_ds.column_names and "speaker_id" not in raw_ds.column_names:
    print("Unsloth: No speaker found, adding default source")
    new_column = ["0"] * len(raw_ds)
    raw_ds = raw_ds.add_column("source", new_column)
elif "source" not in raw_ds.column_names and "speaker_id" in raw_ds.column_names:
    speaker_key = "speaker_id"

# ==================== STEP 6: RESAMPLE AUDIO ====================
target_sampling_rate = 24000
raw_ds = raw_ds.cast_column("audio", Audio(sampling_rate=target_sampling_rate))

print(f"✓ Dataset ready: {len(raw_ds)} samples")
print(f"✓ Speaker column: {speaker_key}\n")

# ==================== STEP 7: OPTIONAL SPLIT ====================
split_ds = raw_ds.train_test_split(test_size=0.1, seed=42)
train_ds = split_ds['train']
val_ds = split_ds['test']

print(f"✓ Train: {len(train_ds)} samples")
print(f"✓ Validation: {len(val_ds)} samples")

# Continue with your Unsloth training!

Key Changes to Avoid Rate Limits:

1. ✅ Pass token=HF_TOKEN explicitly to snapshot_download()
2. ✅ Set max_workers=1 to reduce concurrent requests
3. ✅ Add time.sleep(2) after login to ensure auth propagates
4. ✅ Verify authentication with api.whoami() before downloading
5. ✅ Use resume_download=True to handle interruptions

With Transformers

from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("your-tts-model")

def preprocess_function(examples):
    audio_arrays = [x["array"] for x in examples["audio"]]
    
    inputs = processor(
        text=examples["text"],
        audio=audio_arrays,
        sampling_rate=24000,
        return_tensors="pt",
        padding=True
    )
    
    return inputs

# Apply preprocessing
processed_ds = dataset.map(
    preprocess_function,
    batched=True,
    remove_columns=dataset.column_names
)

Dataset Statistics

• Total Samples: 1,283
• Unique Speakers: 3
• Text Statistics:
  • Average length: 24.0 characters
  • Min length: 4 characters
  • Max length: 141 characters
  • Language: Hausa with proper Unicode diacritics

Data Source

This dataset was created using Mimic Studio recordings for Hausa language TTS development.

Intended Use

This dataset is intended for:

• Training Hausa Text-to-Speech models
• Fine-tuning multilingual TTS models on Hausa
• Research in low-resource language TTS
• Multi-speaker TTS model development

Limitations

• Limited to 3 speakers (may affect speaker diversity in trained models)
• Relatively small dataset size (1,283 samples)
• Audio quality depends on recording conditions

Citation

If you use this dataset, please cite:

@dataset{hausa_tts_embed,
  author = {Aybee5},
  title = {Hausa TTS Dataset (HausaTTSEmbed)},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Aybee5/HausaTTSEmbed}
}

License

Please specify your license here.

Contact

For questions or issues regarding this dataset, please open an issue in the dataset repository.