# cilla-lite

> Gated, non-commercial (CC BY-NC 4.0): ask for access on this page. Cilla Lite is the small 6-layer model for CPUs, distilled from Cilla with guidance baked in (one pass per frame). It speaks Persian in the voice of a short audio prompt. Built on Pocket TTS by Kyutai.

- **Organisation:** mehdi-hf
- **Category:** Text-to-Speech
- **Parameters:** 110M
- **Licence:** CC BY-NC 4.0 (Non-commercial)
- **Released:** 2026-11-14
- **Hugging Face:** 0 downloads (30 days), 68 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q4_K_M | 0.57 |
| Q8_0 | 0.63 |
| F16 | 0.74 |

## About

Gated, non-commercial (CC BY-NC 4.0): ask for access on this page. Cilla Lite is the small 6-layer model for CPUs, distilled from Cilla with guidance baked in (one pass per frame). It speaks Persian in the voice of a short audio prompt. Built on Pocket TTS by Kyutai.

Speed: 1.88x real time on one thread of a cloud x86 CPU (Intel(R) Xeon(R) CPU @ 2.80GHz), measured while making the samples below.

speak_fa.py normalises the text, speaks one sentence at a time from the same prompt, trims the silence each piece opens with and joins them with a short pause. Do the same in your own code: cutting a sentence into 50-token pieces (the old Pocket TTS default) made these models repeat words, once the name «ساندر پیچای» twice in a row.

The config needs the fork. config.yaml sets capitalize_first_letter, append_terminal_punctuation and pad_with_spaces_for_short_inputs to false, which the released pocket-tts does not know and rejects. The loud error is deliberate: without them the package upper-cases or drops the first letters of Persian text.

Write ordinary Persian with punctuation: commas and full stops shape the phrasing. The model reads the ezafe from the context (there are no ezafe marks to type); it still gets some wrong.

The voice is cloned from a prompt: a whole short clip (3 to 10 s) that ends at a pause, in a quiet room. Prompts cut mid-word make the model sound broken. prompts/ has two neutral ones, voice1.wav (a woman) and voice2.wav (a man).

## Alternatives

- [OpenMOSS-Team/MOSS-TTS-Nano-100M](https://selfhost.directory/models/openmoss-team/moss-tts-nano-100m) (100M, Apache 2.0, ≈0.72 GB): MOSS-TTS-Nano is an open-source multilingual tiny speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it is designed for realtime speech generation, can run directly on CPU without a GPU, and keeps the deployment stack simple enough for local demos, web serving…
- [hexgrad/Kokoro-82M](https://selfhost.directory/models/hexgrad/kokoro-82m) (82M, Apache 2.0, ≈0.68 GB): Kokoro is an open-weight TTS model with 82 million parameters. Despite its lightweight architecture, it delivers comparable quality to larger models while being significantly faster and more cost-efficient.
- [zaakirio/kokoro-ru](https://selfhost.directory/models/zaakirio/kokoro-ru) (82M, OpenRAIL, ≈0.68 GB): Russian text-to-speech at 82M parameters, running 9.8x faster than realtime on a laptop CPU with no GPU. Three fixed voices from consented studio actors.
- [KittenML/kitten-tts-2-mini](https://selfhost.directory/models/kittenml/kitten-tts-2-mini) (82M, Apache 2.0, ≈0.68 GB): 82M parameters, 16 languages, 118 voices, 24 kHz. ONNX with int8 weights (82.7 MB; runs on CPU, no PyTorch needed), and the float32 PyTorch weights.
- [Oomni/mithi-v1-flash](https://selfhost.directory/models/oomni/mithi-v1-flash) (99M, Apache 2.0, ≈0.72 GB): A small text-to-speech model for 11 Indian languages, with two voices: Satvik (male) and Charvi (female).
- [hexgrad/Kokoro-82M-v1.1-zh](https://selfhost.directory/models/hexgrad/kokoro-82m-v1.1-zh) (82M, Apache 2.0, ≈0.68 GB): Kokoro is an open-weight series of small but powerful TTS models.

More: https://selfhost.directory/models/mehdi-hf/cilla-lite/alternatives

## Runs well on

- [GeForce RTX 4060](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/rtx-4060): F16
- [Apple M4 (16 GB)](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/apple-m4-16gb): F16
- [Radeon RX 6700 XT](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/rx-6700-xt): F16
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/rtx-3060): F16
- [Arc B580](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/arc-b580): F16
- [Tesla T4](https://selfhost.directory/models/mehdi-hf/cilla-lite/on/tesla-t4): F16

## Similar size

- [hexgrad/Kokoro-82M](https://selfhost.directory/models/hexgrad/kokoro-82m) (82M)
- [hexgrad/Kokoro-82M-v1.1-zh](https://selfhost.directory/models/hexgrad/kokoro-82m-v1.1-zh) (82M)
- [OpenMOSS-Team/MOSS-TTS-Nano-100M](https://selfhost.directory/models/openmoss-team/moss-tts-nano-100m) (100M)
- [KittenML/kitten-tts-2-mini](https://selfhost.directory/models/kittenml/kitten-tts-2-mini) (82M)
- [ResembleAI/chatterbox-nano](https://selfhost.directory/models/resembleai/chatterbox-nano) (110M)
- [ekwek/Soprano-80M](https://selfhost.directory/models/ekwek/soprano-80m) (80M)

## FAQ

**How much VRAM does cilla-lite need?**
About 0.7 GB at FP16 / BF16.

**Can cilla-lite run on the RTX 4060?**
Yes. It fits at FP16 / BF16 (about 0.7 GB). It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.

**Can I use cilla-lite commercially?**
No. It is released under CC BY-NC 4.0. Non-commercial use only, with attribution.

## Links

- huggingface: https://huggingface.co/mehdi-hf/cilla-lite

---
Source: https://selfhost.directory/models/mehdi-hf/cilla-lite