Back to the universe
Python PyTorch CosyVoice 3 TTS Sorani Kurdish Hugging Face AI/ML 🤗 Hugging Face Release

RegaLabs-TTS (CosyVoice 3 Central Kurdish)

State-of-the-art Central Kurdish (Sorani / سۆرانی) text-to-speech and one-shot voice cloning adaptation on Hugging Face, trained on 53 hours of curated speech data.

Live Audio Sample (Sorani Kurdish)
aran_en021.wav

"سڵاو، بەخێربێن بۆ پڕۆژەی RegaLabs-TTS"

"Hello, welcome to the RegaLabs-TTS project"

the deep dive — architecture & implementation

Engineered and published RegaLabs-TTS, a production-grade Central Kurdish (Sorani) acoustic adaptation based on FunAudioLLM CosyVoice 3. Trained on 53 hours of phonetic Kurdish speech data (35-40h male, 13-18h female), achieving flawless male voice cloning, precise prosodic timing, and natural intonation for low-resource Kurdish speech synthesis. Released openly on Hugging Face and GitHub under Apache-2.0 with complete inference scripts and Gradio demo.

RegaLabs-TTS: CosyVoice 3 Central Kurdish (Sorani) Adaptation

RegaLabs-TTS is a high-quality Central Kurdish (Sorani / سۆرانی) text-to-speech and voice cloning adaptation developed at RegaLabs based on CosyVoice 3 (FunAudioLLM/Fun-CosyVoice3-0.5B-2512).

  • Hugging Face Model: RegaLabs/RegaLabs-TTS
  • GitHub Repository: RegaLabs/RegaLabs-TTS
  • Base Architecture: CosyVoice 3 (0.5B Flow Matching + Multi-speaker Conditioning)
  • License: Apache 2.0 (Checkpoints & Inference Code)

📊 Dataset & Training Milestones

  • 53 Total Hours of cleaned, phonetically aligned Sorani Kurdish speech data.
    • Male Speakers: ~35–40 hours.
    • Female Speakers: ~13–18 hours.
  • Flow Adaptation Model: Step 2300 checkpoint (cosyvoice3_sorani_flow_best_step2300.pt) paired with matching YAML architecture definition.
  • One-Shot Voice Cloning: High similarity, prosody accuracy, and natural vocal timbre from single reference audio samples.

🎧 Interactive Audio Sample

  • Prompt: "سڵاو، بەخێربێن بۆ پڕۆژەی RegaLabs-TTS" (Hello, welcome to the RegaLabs-TTS project)
  • Sample File: aran_en021.wav

💻 Quick Installation & CLI Inference

# 1. Install via pip directly from GitHub
pip install git+https://github.com/RegaLabs/RegaLabs-TTS.git

# 2. Clone base engine and model weights
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
cd CosyVoice
pip install -r requirements.txt

git clone https://huggingface.co/RegaLabs/RegaLabs-TTS regalabs-tts-weights

# 3. Synthesize speech in Kurdish
python regalabs-tts-weights/infer.py \
  --text "سڵاو، بەخێربێن بۆ پڕۆژەی RegaLabs-TTS" \
  --prompt-wav regalabs-tts-weights/samples/aran_en021.wav \
  --prompt-text "ئەمە دەنگی نموونەیە" \
  --out output_sorani.wav

📦 Model Artifacts & Architecture

  • cosyvoice3_sorani_flow_best_step2300.pt — Sorani acoustic/flow adaptation model weights.
  • cosyvoice3_sorani_flow_best_step2300.yaml — Matching flow architecture configuration.
  • sorani/frontend.py — Central Kurdish text normalizer and grapheme-to-phoneme engine.
  • infer.py — Ready-to-run Sorani inference CLI.
  • app.py — Gradio Web UI Live Demo script.