High-quality Text-to-Speech system for Turkic languages, with a primary focus on Kazakh.

Turkic-TTS delivers natural-sounding speech synthesis powered by state-of-the-art neural architectures. Built on Grad-TTS with emotional voice synthesis capabilities, this system supports multiple speakers and emotions, making it ideal for applications ranging from voice assistants to audiobook narration.

• 🗣️ Multi-speaker support - Generate speech with different voice profiles (Female and Male speakers)
• 😊 Emotional synthesis - Control emotions: neutral, happy, sad, angry, scared, surprised
• 🎯 High-quality output - Neural vocoder (HiFi-GAN) produces 22.05kHz audio
• 🔧 Flexible architecture - Based on Grad-TTS with diffusion probabilistic modeling
• 🌍 Turkic language support - Optimized for Kazakh with IPA phonetic conversion for multiple Turkic languages
• ⚡ Fast inference - Adjustable timesteps for speed-quality trade-off

The system consists of three main components:

1. Text Encoder - Converts text to phoneme embeddings with speaker/emotion conditioning
2. Diffusion Decoder - Grad-TTS based mel-spectrogram generator with emotion control
3. Neural Vocoder - HiFi-GAN converts mel-spectrograms to high-fidelity audio

• Monotonic Alignment Search (MAS) for text-audio alignment
• Classifier-free guidance for enhanced emotion control
• Exponential Moving Average (EMA) training for stability
• Support for both emotional and neutral speech synthesis

• Python 3.9 or higher
• PyTorch 1.10+
• CUDA (optional, for GPU acceleration)

1. Clone the repository

git clone https://github.com/stukenov/turkic-tts.git
cd turkic-tts

2. Install dependencies

pip install -r requirements.txt

3. Build monotonic alignment module

cd model/monotonic_align
python setup.py build_ext --inplace
cd ../..

Pre-trained models are available on HuggingFace:

• Main repository: stukenov/turkic-tts-models

Download the models and place them in the appropriate directories:

• TTS model checkpoint → pt_10000/
• HiFi-GAN vocoder → pre_trained_3/

Create a text file with your input (e.g., filelists/my_text.txt):

Сәлем, қалайсың?|0|1

Format: text|emotion_id|speaker_id

Emotion IDs:

• 0: Angry
• 1: Fear
• 2: Happy
• 3: Neutral
• 4: Sad
• 5: Surprised

Speaker IDs:

• 0: M1 (Male 1)
• 1: F1 (Female 1)
• 2: M2 (Male 2)

Run inference:

python inference_EMA.py \
  -c configs/train_grad.json \
  -m pt_10000/EMA_grad_10000.pt \
  -t 10 \
  -g 100 \
  -f filelists/my_text.txt \
  -r output_audio/

Parameters:

• -c: Configuration file
• -m: Model checkpoint path
• -t: Number of diffusion timesteps (higher = better quality, slower)
• -g: Classifier-free guidance level (recommended: 100)
• -f: Input text file
• -r: Output directory for audio files

# Coming soon: Python API example

1. Organize your dataset with audio files and transcriptions
2. Create filelists in the format: audio_path|speaker_id|emotion_id|text

python data_preparation.py -d /path/to/your/dataset

CUDA_VISIBLE_DEVICES=0 python train_EMA.py \
  -c configs/train_grad.json \
  -m logs/train_logs

The training script includes:

• Exponential Moving Average (EMA) for model stability
• Duration prediction with Monotonic Alignment Search
• Multi-speaker and emotion embedding
• Tensorboard logging

The model has been trained on high-quality Kazakh speech data with:

• Multiple speakers and emotional expressions
• 80-dimensional mel-spectrograms
• 22.05kHz sampling rate
• Diffusion-based generation with controllable quality

Key configuration options in configs/train_grad.json:

• Model architecture: encoder layers, channels, attention heads
• Training parameters: learning rate, batch size, optimization
• Data processing: sampling rate, hop length, mel bins
• Speaker/emotion settings: number of speakers, emotions, embedding dimensions

turkic-tts/
├── model/               # Core model architectures
│   ├── tts.py          # Main Grad-TTS model
│   ├── diffusion.py    # Diffusion decoder
│   ├── text_encoder.py # Text encoding module
│   └── monotonic_align/ # MAS alignment
├── text/               # Text processing
│   ├── cleaners.py     # Text normalization
│   └── symbols.py      # Phoneme symbols
├── configs/            # Configuration files
├── inference_EMA.py    # Inference script
├── train_EMA.py        # Training script
├── models.py           # HiFi-GAN vocoder
├── ipa_convert.py      # IPA phonetic conversion
└── requirements.txt    # Dependencies

While optimized for Kazakh, the IPA conversion module includes support for:

• Kazakh
• Turkish
• Kyrgyz
• Uzbek
• Azerbaijani
• Turkmen
• Tatar
• Bashkir
• Sakha (Yakut)
• Uyghur

Contributions are welcome! Whether it's:

• Bug fixes
• New features
• Documentation improvements
• Training data contributions
• Support for additional Turkic languages

Please feel free to submit issues and pull requests.

If you use this code or models in your research, please cite:

@software{turkic_tts_2024,
  author = {Tukenov, Saken},
  title = {Turkic-TTS: High-Quality Text-to-Speech for Turkic Languages},
  year = {2024},
  url = {https://github.com/stukenov/turkic-tts}
}

This project is licensed under the MIT License - see the LICENSE file for details.

This work builds upon several excellent open-source projects:

• Grad-TTS - Diffusion probabilistic TTS
• HiFi-GAN - Neural vocoder
• KazEmoTTS - Kazakh emotional TTS dataset

For questions, suggestions, or collaborations:

• GitHub: @stukenov
• Issues: Project Issues

Star ⭐ this repository if you find it useful!