servinosmanov/tts-crh-sevil-fixed

Crimean Tatar TTS Dataset - Sevil (Female Voice) - Fixed Version

This is a fixed version of the speech-uk/tts-crh-sevil dataset.

Improved Crimean Tatar TTS Dataset (Sevil Speaker)
Author: Servin Osmanov
Dataset URL: https://huggingface.co/datasets/servinosmanov/tts-crh-sevil-fixed

---

🧩 Dataset Summary

tts-crh-sevil-fixed is an improved and fully cleaned version of the original Crimean Tatar TTS dataset featuring the “Sevil” female speaker.
This dataset was reconstructed, corrected, and enhanced by Servin Osmanov as part of a long-term effort to build high-quality AI and TTS tools for the Crimean Tatar (crh) language — one of the world’s severely under-resourced Turkic languages.

The dataset resolves audio issues, transcript errors, and metadata inconsistencies present in the original dataset, making it suitable for training modern high-accuracy TTS models such as XTTS-v2, VITS, Glow-TTS, and Tacotron-based architectures.

A detailed walkthrough of the improvements is available here:
👉 https://dev.to/servin_osmanov/my-journey-improving-a-tts-model-for-the-crimean-tatar-language-53f7

---

🎯 Motivation

High-quality TTS datasets for Crimean Tatar are extremely rare.
Existing datasets suffer from:

• noisy audio
• inconsistent text
• missing or incorrect metadata
• misaligned text–audio pairs

This improved dataset was created to:

• provide a reliable foundation for TTS model training
• repair and enhance the widely-used Sevil dataset
• support the development of AI tools for Crimean Tatar
• enable research on Turkic low-resource speech technologies
• allow practitioners to train XTTS-v2 and similar models in a clean environment

---

✨ Key Improvements Over Previous Dataset

✔ 1. Audio Quality Fixes

• removed background noise
• normalized volume
• trimmed excessive silence
• removed duplicate and corrupted audio files

✔ 2. Transcript Corrections

• fixed transcription errors
• restored missing text
• aligned text with audio
• removed broken characters and encoding issues

✔ 3. Metadata Reconstruction

• rewritten metadata.csv
• standardized file names
• added normalized text
• ensured consistent formatting for training pipelines

✔ 4. XTTS-v2 Ready

• validated with Coqui TTS tools
• cleaned text fits XTTS language rules
• verified through actual model training
• proven to significantly improve output quality

✔ 5. Used in Real TTS Model Training

This dataset was successfully used to train improved Crimean Tatar voice models, demonstrating:

• better pronunciation
• fewer artifacts
• more stable attention alignment
• clearer prosody

Key Improvements

• ✅ 100% working audio files (1566/1566)
• ✅ Fixed audio corruption - all WAV files re-encoded with ffmpeg
• ✅ Verified quality - all files tested and validated
• ✅ Better model training - loss improved from 0.276 → 0.267

Original vs Fixed

Metric	Original Dataset	Fixed Dataset
Total files	1,566	1,566
Working files	415 (26.5%)	1,566 (100%)
Model test success	47% (on 415 files)	100% (on 1,566 files)
Training loss	0.276	0.267

---

🏗 Dataset Structure

tts-crh-sevil-fixed/

│── audio/ # WAV files, 16kHz mono

│── metadata.csv # file_name | text | text_cleaned

│── README.md

---

📋 Data Fields

• file_name — audio file (WAV)
• text — original transcript
• text_cleaned — normalized and cleaned text (ready for TTS training)

---

🗣 Language Information

• Language: Crimean Tatar (ISO: crh)
• Script: Latin
• Speaker: Single female speaker (Sevil)
• Domain: General, conversational

---

🔧 Preprocessing Pipeline

Audio Processing:

• resampled to 16kHz
• converted to mono
• peak normalization
• silence trimming
• noise reduction
• removal of defective files

Text Processing:

• unicode normalization
• consistent punctuation rules
• removal of extraneous characters
• handcrafted fixes for alignment

---

📚 Applications

This dataset is designed for:

• TTS model training (XTTS-v2, VITS, Glow-TTS, Tacotron2)
• voice cloning research
• ASR experiments for low-resource languages
• linguistic analysis of Turkic languages
• educational and cultural preservation AI tools

Currently used in:

• Ana-Yurt educational apps
• Crimean Tatar dictionary tools
• Research on low-resource TTS model improvements

---

⚖ Ethical Considerations

• Single-speaker dataset; voice cloning misuse is discouraged.
• Do not use for political deepfake applications.
• Allowed for research, education, and community-supporting projects.

---

📄 License

CC-BY-4.0

---

🧪 Usage Example

from datasets import load_dataset

ds = load_dataset("servinosmanov/tts-crh-sevil-fixed")
print(ds["train"][0])

Training a TTS Model

from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
from datasets import load_dataset

# Load dataset
dataset = load_dataset("servinosmanov/tts-crh-sevil-fixed")

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

# Prepare dataset
def prepare_dataset(example):
    audio = example["audio"]
    example = processor(
        text=example["transcription"],
        audio_target=audio["array"],
        sampling_rate=audio["sampling_rate"],
        return_attention_mask=False,
    )
    return example

dataset = dataset.map(prepare_dataset, remove_columns=dataset["train"].column_names)

# Train model...

Demo

crh-tts-sevil-demo

🙏 Acknowledgements

Special thanks to:

• the Crimean Tatar community.
• Original dataset: speech-uk/tts-crh-sevil
• contributors of the original Sevil dataset
• Ana-Yurt, Qırım Junior
• researchers and developers working on low-resource AI technologies

🔖 Citation

If you use this dataset in research or development, please cite:

Osmanov, Servin (2025). tts-crh-sevil-fixed: Improved Crimean Tatar TTS Dataset. HuggingFace.
https://huggingface.co/datasets/servinosmanov/tts-crh-sevil-fixed