akalandia/ka-geo-voice-male-v1

Dataset Card for Georgian Male Voice Dataset v1

Intended Use

• Primary Use: Training and fine-tuning TTS models for Georgian language synthesis, including microsoft/speecht5_tts.
• Secondary Use: Research in speech synthesis, voice conversion, or linguistic analysis.

SpeechT5 Compatibility

This dataset is specifically formatted to be compatible with microsoft/speecht5_tts fine-tuning. The dataset includes:

• Audio: 22,050 Hz mono WAV files (matching SpeechT5 requirements)
• Sentence: Georgian text transcriptions
• Speaker ID: Speaker identifier (0 for this single-speaker dataset)

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/ka-geo-voice-male-v1")

# Access examples
print(dataset["train"][0])
# Output: {'audio': {...}, 'sentence': 'გამარჯობა, როგორ ხარ?', 'speaker_id': 0}

Fine-tuning SpeechT5 with this Dataset

from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
from datasets import load_dataset

# Load processor and model
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")

# Load dataset
dataset = load_dataset("your-username/ka-geo-voice-male-v1")

# The dataset is ready for SpeechT5 fine-tuning
# Follow the standard SpeechT5 fine-tuning procedure

Data Collection

• Source: Audio recordings from a native Georgian male speaker.
• Recording Setup: [Details not provided; assume standard recording conditions]
• Preprocessing: Audio normalized using FFmpeg with loudness normalization (loudnorm) and silence removal (silenceremove).

Data Structure

• Format: WAV
• Sample Rate: 22,050 Hz
• Channels: 1 (mono)
• Directory Structure:

  ka-geo-voice-male-v1/
  ├── README.md
  ├── ka-geo-voice-male-v1.py  # Dataset loading script
  ├── metadata.csv
  └── wav/
      ├── 000001.wav
      ├── 000002.wav
      ├── 000003.wav
      ├── 000004.wav
      └── 000005.wav
  

• Metadata File: metadata.csv contains comma-separated values with columns: file_name, sentence, speaker_id. Example:

  file_name,sentence,speaker_id
  wav/000001.wav,"გამარჯობა, როგორ ხარ?",0
  wav/000002.wav,"დღეს ამინდი ძალიან კარგია.",0
  

  Note: Sentence values are quoted to handle commas within the text.

Dataset Fields

When loaded via the Hugging Face datasets library, each example contains:

• audio: Audio object with the waveform array and sampling rate (22,050 Hz)
• sentence: String containing the Georgian text transcription
• speaker_id: Integer identifier for the speaker (0 for all samples in this single-speaker dataset)

Annotations

• Type: Transcripts
• Language: Georgian
• Quality: Manually verified transcripts corresponding to the audio content.

Dataset Statistics

• Number of Samples: 5
• Total Duration: Approximately 23.2 seconds
• Average Sample Duration: ~4.6 seconds
• Speaker Demographics: Male, native Georgian speaker
• Audio Quality: 22,050 Hz, mono, 16-bit WAV

Requirements

To use this dataset, you'll need:

pip install datasets transformers soundfile librosa

For SpeechT5 fine-tuning specifically:

pip install transformers[torch] datasets soundfile librosa accelerate

Ethical Considerations

• Ensure the dataset is used responsibly, respecting privacy and cultural sensitivities.
• The speaker consented to the recordings for research purposes.

Citation

If you use this dataset in your research, please cite it as:

@dataset{ka_geo_voice_male_v1_2025,
  title={Georgian Male Voice Dataset v1},
  author={Georgian TTS Contributors},
  year={2025},
  publisher={Hugging Face},
  howpublished={Hugging Face Datasets},
  url={https://huggingface.co/datasets/ka-geo-voice-male-v1}
}