Silero Models

Local text-to-speech models generate multilingual audio on CPU or GPU through PyTorch, with SSML controls and MIT-licensed CIS base models.

Silero Models is a collection of pretrained text-to-speech models for developers adding voice output to applications or services. It suits projects that need speech generation on their own hardware, including support for Turkic and Caucasian languages.

Languages and voices: Generate speech in Russian, Uzbek, Kyrgyz, Tatar, Chechen, Avar and other supported languages, with multiple speaker voices available for some languages.
Russian pronunciation: Russian-only models can assign word stress automatically and distinguish words with identical spellings but different pronunciations; some also handle question intonation.
Speech controls: SSML support lets applications control speech delivery across the Russian, Turkic, Caucasian and CIS base model families.

The models run through PyTorch on CPU or GPU. CIS base models require a PyTorch-compatible system and an AVX2-capable processor on x86/64 hardware.

Capabilities vary by model family. Turkic and Caucasian models don't support stress handling. CIS base models lack automatic stress and homograph handling: the standard variant requires stress-marked words in every language, while the nostress variant requires those markings only for Russian, Belarusian and Ukrainian. The CIS base models use the MIT license.

Text-to-Speech Models Text-to-Speech and Audiobooks Multilingual Works offline