Cilla Mini is a Persian speech synthesis model trained by community author mehdi-hf based on Kyutai's Pocket TTS. It uses 1820 million parameters during generation, outputs 24 kHz mono audio, and clones the voice from a reference audio clip of 3–10 seconds. The author reports approximately 3.79x real-time speed on a single-threaded CPU on a cloud server, and about 12–23x real-time on an M1 Pro using the Phonon runtime written in Rust; the Hugging Face demo space runs directly in the browser via WebAssembly, with no text uploaded. The same series includes Cilla Lite, a 1 billion parameter model designed for CPUs, which the author claims has a lower word error rate (6.7% vs. 10.4%).
Why it matters Open-source Persian speech synthesis is rare; it is small enough to run offline in a web browser, keeping text on the local device, making it suitable for screen readers and offline assistants, but it is restricted to non-commercial use.
Hugging Face likes 22 · Hugging Face trending score 21 · Hugging Face downloads 0
License: Non-commercial · ComfyUI: None · VRAM: ≈1 GB estimated
Reasons
- Includes an online browser demo, Mac runtime scripts, and a JavaScript library
- The model card clearly documents evaluation methods, test conditions, and limitations
Usage
- After requesting access and logging in on the Hugging Face page, run pip install "pocket-tts>=4.0.1", hf download mehdi-hf/cilla-mini --exclude "model.safetensors" "phonon/*" --local-dir cilla-mini, enter the directory and run python speak_fa.py --voice prompts/parisa.wav --text "Persian text" --out hello.wav; you can also directly open the demo space mehdi-hf/cilla-mini-demo
Risks
- CC BY-NC 4.0, non-commercial use only; supports only Persian, Latin alphabet words and emojis are directly removed, and Chinese is not supported
- Long texts must be synthesized sentence by sentence, otherwise repetition or missing words may occur; consent must be obtained from the individual when cloning voices, and generated speech should be labeled as synthetic