Silencio-2 released as an open-weight emotional TTS model
The author released Silencio-2, an 18.2M-parameter text-to-speech model fine-tuned from Kyutai's Pocket TTS, with 24 kHz streaming, voice cloning and inline emotion and sound tags. Weights and code are available on Hugging Face, while a browser-based Colab demo is also provided. The author says it runs at roughly 20 times realtime on an Apple M4 CPU, but clarifies that this came from one long-text laptop run rather than a benchmark; no WER benchmark was run on the released checkpoint.
Silencio-2 adds a compact, open-weight TTS option with expressive tags, local deployment and voice cloning, although its performance and tag consistency have not been independently verified.
I made this, so treat it as biased. It is free and open, with no paid tier or API behind it.
Silencio-2 is a fine-tune of Kyutai's Pocket TTS 18M that follows tags written inside the text:
What a long day. [sigh] <whisper> Keep your voice down. <whisper> Keep your voice down. <angry> I said stop! Weights and code: https://huggingface.co/BlueRoseLabs/Silencio-2
Try it in your browser, no setup: https://colab.research.google.com/drive/1d6GSiRgXbYyBo91DCMrSr5y2LR_peJKy?usp=sharing
The numbers
#Silencio-2#BlueRose Labs#Kyutai#Pocket TTS#Hugging Face#AudioSet#LibriSpeech