You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Cilla Mini: Persian text to speech (18.2M)

Gated, non-commercial (CC BY-NC 4.0): ask for access on this page. Cilla Mini is the 18M model for CPUs, phones and browsers: Cilla shrunk with Kyutai's small-model recipe (8 layers, a 64-dim interface, a one-step drifting sampler head, the 2.6M Mimi decoder), guidance baked in. It speaks Persian in the voice of a short audio prompt. Built on Pocket TTS by Kyutai.

Try it in your browser · Run it on a Mac · Phonon runtime · Browser library · Python package

parameters used for generation 18.2M = 15.6M flow language model (8 layers, with the sampler head) + 2.6M Mimi decoder. The file holds 27.9M: the other 9.7M is the Mimi encoder, which runs once per voice prompt
sample rate 24 kHz, mono
language Persian (fa)
weights model.safetensors, 107 MB (the Pocket TTS format); phonon/model/model.safetensors is the same model laid out for the Phonon runtime
sibling mehdi-hf/cilla-lite: Cilla Lite, 100M

Speed

Audio made per second of work ("real time" is 1x):

Where Speed
Python (pocket-tts), one thread of a cloud x86 CPU (Intel(R) Xeon(R) CPU @ 2.80GHz) 3.79x
Phonon runtime ptts, one core of an Apple M1 Pro about 12 to 23x (first audio in 30 to 95 ms)
In a browser tab (WebAssembly), Chrome on an M1 Pro about 9 to 17x (short texts are at the slow end)

Use

Python (the pocket-tts package)

pip install "pocket-tts>=4.0.1"                                                     # the released package
hf auth login                                                                  # the repo is gated: ask for access on its page first
hf download mehdi-hf/cilla-mini --exclude "model.safetensors" "phonon/*" --local-dir cilla-mini && cd cilla-mini   # the code, prompts and config
python speak_fa.py --voice prompts/parisa.wav --text "سلام، حال شما چطور است؟" --out hello.wav   # fetches the weights once

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 released package is enough (pocket-tts 4.0.1 or newer, checked for this repo). config.yaml sets capitalize_first_letter, append_terminal_punctuation and pad_with_spaces_for_short_inputs to false: leave them, otherwise the package upper-cases or drops the first letters of Persian text.

On a Mac, fast: the Phonon runtime

The phonon/ folder has the model for the Phonon runtime (Rust, no PyTorch, no GPU) and the same Persian front-end. Apple silicon:

hf download mehdi-hf/cilla-mini --include "phonon/*" --local-dir cilla-mini
cd cilla-mini/phonon
bash setup.sh        # once: downloads the official runtime and the browser library, checks their checksums
bash try_it.sh       # four demo clips into out/
python3 say_fa.py --voice parisa --text "سلام، این یک آزمایش است." --out out/mine.wav --play
bash run_web.sh      # a web page with captions: http://127.0.0.1:8777

Details, options and notes: phonon/README.md.

In a browser

The demo Space mehdi-hf/cilla-mini-demo runs the model inside the tab (WebAssembly, the phonon-tts library): the text is not sent anywhere, and the page asks before it downloads the 107 MB, which the browser then keeps. Open it in its own tab for the multi-threaded mode. Tick "Stream while generating" and the first sound plays about 20 ms after Speak, each 80 ms chunk the moment it is made. The same page runs from phonon/run_web.sh, and phonon/web/persian.js is the JavaScript port of the Persian text steps for your own web app.

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.

Voices

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 parisa.wav, the voice of Parisa (recorded by her and used here with her agreement), and two neutral ones, voice1.wav (a woman) and voice2.wav (a man).

In the check, every clip made from each of them was closest to that prompt, so the model answers in the voice you give it.

Samples

Made with speak_fa.py: Parisa's voice is prompts/parisa.wav, the female voice is prompts/voice1.wav, the male voice is prompts/voice2.wav.

News

Voice Sample
Parisa
Female
Male

بر اساس گزارش سازمان هواشناسی، بارش باران و برف در بیشتر استان‌های شمالی و غربی کشور تا پایان هفته ادامه خواهد داشت. از رانندگان خواسته شده است که در جاده‌های کوهستانی با احتیاط بیشتری حرکت کنند و زنجیر چرخ همراه داشته باشند.

Story

Voice Sample
Parisa
Female
Male

پیرمرد از پنجره‌ی کوچک اتاق به کوچه نگاه کرد. باران ریزی می‌بارید و چراغ‌های زرد خیابان روی آسفالت خیس می‌درخشیدند. آهی کشید و با خودش گفت که این شهر دیگر شهر روزهای جوانی‌اش نیست.

Names and technology

Voice Sample
Parisa
Female
Male

شرکت گوگل در ماه گذشته از نسخه‌ی تازه‌ای از مدل هوش مصنوعی خود رونمایی کرد. مدیرعامل این شرکت، ساندر پیچای، در سخنرانی خود در شهر سانفرانسیسکو گفت که این مدل می‌تواند متن، تصویر و صدا را با هم بفهمد و در بیش از صد زبان، از جمله فارسی، پاسخ دهد.

Evaluation

Recognition by a Persian ASR on real recordings, voices cloned from other clips of the same speaker (step 100,000, EMA; the guidance is built in). The ASR itself scores 8.6% word error on the real recordings, so that is the floor.

real recordings, WER 10.4%
real recordings, CER 3.9%
speaker similarity 0.953

These measure intelligibility and a coarse speaker match; they do not show whether a particular voice is reproduced (see Voices) or how often the ezafe is right, which needs listening.

Intended use

Speech on the user's own device: offline and private assistants, screen readers and other accessibility tools, games, and apps that cannot wait for a cloud round trip. Run on a Mac or in a browser tab, nothing leaves the device.

Limitations

  • Persian only. Ezafe mistakes happen, mostly on rare words and in long noun phrases.
  • Only Persian script is spoken: speak_fa.py drops English (Latin-script) words and other foreign characters, such as emoji, from the text without a warning, so write names and terms in Persian letters (for example «گوگل» for Google). Digits are read as Persian words.
  • Voice cloning is not exact; it follows the prompt's timbre and style, not every detail.
  • Long text must be spoken sentence by sentence (above), or words are repeated or lost.
  • The Phonon files (phonon/) hold three voices, and the runtime's own voice cloning cannot be used with this model; a new voice needs the Python tools.
  • Speech made with it should be labelled as synthetic; do not clone a voice without the person's consent.

Licence and credits

CC BY-NC 4.0: non-commercial use only. Commercial use needs a separate licence: mehdi@oncillaai.com.

Built on Pocket TTS by Kyutai (kyutai-labs/pocket-tts); the Mimi codec and the 2.6M decoder this model uses are Kyutai's.

Downloads last month
-
Safetensors
Model size
27.9M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support