nibor1896/horus.f5tts.onnx

This is a .NET library for cloning a voice from a short audio clip and using it to speak new text, based on an AI model called F5-TTS. Normally running F5-TTS requires Python and PyTorch, but this library runs the same model using ONNX Runtime, a way of running trained AI models directly, so a C# or .NET application can use it without any Python installed. It can run on a regular CPU or on a GPU, including AMD, Intel, and Nvidia graphics cards through DirectX or CUDA. To use it, you give the library a short reference recording of a voice along with a transcript of what is said in that clip, plus the new text you want spoken. It returns 24 kHz audio in that cloned voice. The model files themselves are not included in the package and need to be downloaded separately, either from ready-made German and English exports the author provides, or exported yourself for other languages using a separate tool. The voice you hear comes from the reference clip you provide, while the language and accent come from which model checkpoint you load, so an English checkpoint given German text will not sound right. The library includes handling for longer pieces of text, which the underlying model cannot generate well in a single pass past about 22 seconds combined with the reference clip. It automatically splits long text into sentence-sized pieces, generates each one, and stitches them together smoothly. There is also a streaming mode that returns each sentence's audio as soon as it is ready, so playback can start after the first sentence instead of waiting for the whole paragraph, while still producing audio identical to the non-streaming version. The library itself does not play audio, since it is meant to stay lightweight, so a separate audio library like NAudio is needed to actually hear the output, though raw audio samples are also available for feeding into any custom audio pipeline.