🎙️ WhisperSpeech2

An Open Source text-to-speech system built by inverting Whisper. This is a fork of WhisperSpeech optimized for inference that introduces ⚡"cuda graphs"⚡ for faster inference.

🚀 Installation

pip install whisperspeech2

PyTorch

CUDA libraries if using an Nvidia GPU (tested with torch 2.9 on CUDA 12.8 and torch 2.14.1 on CUDA 13.0).

✨ Available Models

You can mix and match the models for different quality and compute requirements. See the WhisperSpeech Hugging Face repository.

S2A Models (Semantic to Acoustic)

T2S Models (Text to Semantic)

v1.1 Small and HQ Fast never ended a sentence early in testing, while the other T2S models occasionally do.

Benchmark (no cuda graph)

Benchmark (with cuda graph)

People with Nvidia GPUs can set the "use_cuda_graph" parameter to "true" for faster processing.

Examples

See the examples/ directory for more usage examples including GUI applications and streaming playback.

Thanks/Shout Outs

Thanks to Jakub for the inspiration.

The fine folks at the Dia2 for forcing me to learn about cuda graph. Go check theirs out too.

BIG FUCK YOU

And finally a big "fuck you" to Microsoft for installing shit on my computer that I don't need and/or want for years.

Your pathetic VibeVoice project just got pwned!

You can end support for Windows 10 but still find ways to surreptitiously install Microsoft Edge, the shittest browser ever, on my computer without notification or approval whatsoever...or consistently defeat ways to simpy install Windows without having to register an account online...again, a big FUCK YOU to Microsoft!