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

RegaLabs-TTS is a high-quality Central Kurdish (Sorani / سۆرانی) text-to-speech adaptation developed by RegaLabs based on CosyVoice 3.

• Hugging Face Model: RegaLabs/RegaLabs-TTS
• GitHub Repository: RegaLabs/RegaLabs-TTS

---

---

🎧 Audio Example

Listen to a generated Sorani audio sample:

🔊 Download Sample Audio WAV

---

📊 Dataset & Model Capabilities

• Dataset Volume: Trained on 53 hours of high-quality Sorani Kurdish speech data.
  • Male Speakers: ~35–40 hours.
  • Female Speakers: ~13–18 hours.
• Voice Cloning Performance:
  • Male Voices: Evaluated and tested — clones male voices flawlessly with high similarity, prosody accuracy, and natural tone.
  • Female Voices: Untested — zero-shot female voice cloning has not been evaluated yet.

---

💻 Installation & Usage

1. Install via pip

pip install git+https://github.com/RegaLabs/RegaLabs-TTS.git

2. Download Model Weights & Run Inference

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

# 2. Clone RegaLabs-TTS model repo
git clone https://huggingface.co/RegaLabs/RegaLabs-TTS regalabs-tts-weights

# 3. Synthesize speech (the Sorani LLM adapter and flow checkpoint ship in the repo)
python regalabs-tts-weights/infer.py \
  --text "سڵاو، بەخێربێن بۆ پڕۆژەی RegaLabs-TTS" \
  --prompt-wav regalabs-tts-weights/samples/aran_en021.wav \
  --prompt-text "دەنگێکی لەسەرخۆ، هێمن و پڕ لە بڕوابەخۆبوون." \
  --out output_sorani.wav

3. Launch Web Demo (Gradio)

python regalabs-tts-weights/app.py

---

📦 Model Artifacts

• cosyvoice3_sorani_flow_best_step2300.pt — Sorani acoustic/flow adaptation model weights (Step 2300).
• cosyvoice3_sorani_lora_refined_best.pt — Sorani LLM LoRA adapter (required for Sorani speech).
• cosyvoice3_sorani_flow_best_step2300.yaml — Matching flow architecture configuration.
• infer.py — Ready-to-run Sorani inference script.
• app.py — Gradio Web UI Live Demo script.
• sorani/censor.py — Fail-closed Sorani sexual-word filter with integrity checks (see below).
• samples/aran_en021.wav — Sample generated audio snippet.
• assets/ — Project banners & graphics.

---

🚫 Content Filtering & Model Integrity

RegaLabs-TTS ships with a fail-closed Sorani content filter (sorani/censor.py). Sexual obscenities are bleeped (......) before synthesis instead of being spoken.

• Blocked vocabulary: a curated, dictionary-verified blocklist of Sorani sexual obscenities (the exact terms are defined in sorani/censor.py). Inflected forms and spelling variants are caught automatically.
• Tamper protection: the word list is SHA-256-signed, and the flow checkpoint must match the official SHA-256 (033abd6f...). If either is modified, synthesis refuses to run — the model "breaks itself" rather than speaking uncensored.
• Re-signing (only for official model updates): python sorani/censor.py --rehash re-signs the word list; python sorani/censor.py --sign-checkpoint PATH signs a newly released official checkpoint.
• Honest limitation: TTS censorship lives in the text layer (the weights themselves cannot refuse words), so a determined attacker with full code access can patch the checks out. This protects against accidental or naive removal and model swaps, not against deliberate reverse engineering.

---

⚠️ Responsible Use & No-Liability Disclaimer

RegaLabs-TTS is a voice-cloning-capable TTS system. It generates whatever text it is given, in whichever voice it is prompted with. RegaLabs:

• is NOT responsible for any content generated with this system, or for how it is used, modified, or redistributed — the user is solely responsible for the legality and consequences of their use;
• prohibits cloning the voice of any real person without that person's explicit consent, and prohibits use for impersonation, fraud, deepfakes, scams, defamation, harassment, sexual content involving minors, or any illegal activity (see Section 4 of the LICENSE);
• warns that all generated audio should be treated as potentially synthetic; never rely on model output as evidence of a real person's words.

Voice consent rule: zero-shot voice cloning is only permitted with the speaker's own approval, or for voices you own or are authorized to use.

---

📜 License & Mandatory Attribution

• Model Checkpoint & Codebase: Licensed under Apache 2.0 by RegaLabs. Commercial and non-commercial use is fully allowed, provided mandatory credit for RegaLabs is included and the consent/prohibited-uses conditions in the LICENSE are respected.
• Stock Voices & Audio Samples: Non-Commercial Use Only. Pre-packaged stock prompt voice samples and demo audio files (including samples in samples/) are strictly restricted to non-commercial use and prohibited for commercial voice cloning/redistribution.

📌 Credit & Attribution Guidelines

Any public use, generated media (videos, podcasts, audiobooks, broadcasts, or AI services), software, or derivative works utilizing RegaLabs-TTS MUST explicitly provide visible credit to RegaLabs:

• For Videos & Audio Content: Include in description/credits: "Voice synthesized using RegaLabs-TTS by RegaLabs" or "Audio powered by RegaLabs".
• For Applications & Software: Include attribution in application credits or about section.