Silencio-2

Silencio-2

Turns English text into 24 kHz speech with emotion and sound tags, and can copy a voice from a short reference clip. A tiny text to speech model fine-tuned from Kyutai's Pocket TTS.

Made by BlueRoseLabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Silencio-2 is a small English text to speech model from BlueRoseLabs. It reads text aloud at 24 kHz in a bundled voice or in a voice copied from a few seconds of reference audio. You steer it inside the text with tags, such as a sigh or a whisper. It is a fine-tune of Kyutai's Pocket TTS.

Who it suits: Developers, game makers and content creators who want expressive English speech from a model that runs on a normal computer.

Industry | Who uses it | How they use it | Example
Software | Developers | Add local speech output to an app | A narrator that reads status messages
Games | Game makers | Voice characters with emotions and sounds | A line spoken in a whisper
Media | Content creators | Draft voice-overs with sighs and laughs in the right places | A short clip with a laugh
Accessibility | Tool builders | Read text aloud with a small, fast model | A screen reader prototype

If you need languages other than English, you can skip it, since the card lists English only.

Where to use it:

• Windows, macOS or Linux with Python
• Python programs that import the silencio2 package from the repository
• CPU-only machines, where the card reports faster than real-time speech

Use case | Who | What they get back
Plain text to speech | Developers | A wav file in one of the bundled voices
Tagged speech | Creators | Sighs, laughs and emotions placed where you write them
Voice cloning | Creators with permission | Speech in a voice taken from a short clean clip
Streaming | App builders | Audio chunks played as they are generated

How to set it up:

• Python, to run the Hugging Face download tool
• Room for a single 112 MB weights file
• The requirements file from the repository, installed from a clone or with the snapshot folder on the Python path

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Install the packages in requirements.txt.
4. Import Silencio2TTS in Python and load the model.

pip install -U huggingface_hub
hf download BlueRoseLabs/Silencio-2
pip install -r requirements.txt

How to use it: Load the model once, then call save to write a wav file. Sounds go directly into the text, an emotion tag applies to the text after it, and a voice can be a bundled name or a clip of your own.

from silencio2 import Silencio2TTS

tts = Silencio2TTS.from_pretrained()
tts.save("sigh.wav", "What a long day. [sigh]")
tts.save("mixed.wav", "<whisper> Keep your voice down. <angry> I said stop!")
tts.save("clone.wav", "Hi! This is my voice.", voice="my_clip.wav")

Tool | What it does
save | Writes the speech to a wav file
generate | Returns the audio as a tensor, with a seed for repeatable takes
stream | Yields audio chunks as they are produced
sounds | Tags for a laugh, sigh, gasp, cough or throat clear
emotions | Tags for happy, sad, angry, fear, surprised or whisper

Example: You save the line "<whisper> Keep your voice down" with the default bundled voice. The model writes a wav file in which that sentence is spoken in a whisper. If you write a tag the model does not know, it stops with an error and does not speak it.

For newcomers and experts: Start with one sentence and one bundled voice, then add a tag such as a sigh and listen to the change. Try voice cloning only with a clean clip of a voice you have the right to use.

Under the hood it is a flow-based language model. For each audio frame it predicts Mimi codec latents, with the text and a voice prompt as input. Cloning relies on an encoder that takes up the rest of the 27.9M parameters, while about 18.2M handle the speech itself. A tokenizer with 11 extra tokens covers the tags.

Write text with tags
Tokenizer reads text and tags
Model predicts Mimi codec latents
Decoder outputs 24 kHz audio

Good to know:

• License: CC BY 4.0
• Cost to run: none named
• Data it reaches: downloads files from Hugging Face, then runs locally and reads any reference clip you give it
• Not for: languages other than English
• Not for: cloning a voice you have no right to use, so use only voices you have permission for

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: pytorch
• Parameters: 27.9M
• Weights: about 112 MB in F32; about 14 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 351
• Likes: 22
• License: cc-by-4.0
• Languages: en
• Based on: kyutai/pocket-tts
• Tags: pytorch, safetensors, text-to-speech, voice-cloning, emotional-speech, streaming, on-device

Links

• Model: https://huggingface.co/BlueRoseLabs/Silencio-2
• Files: https://huggingface.co/BlueRoseLabs/Silencio-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BlueRoseLabs/Silencio-2

Then follow the usage in its model card. Nothing in this kit runs by itself.