audio.cpp 0.3 adds Supertonic 3, MOSS-TTS, IndexTTS2, and Irodori-TTS with GGUF support

The audio.cpp project has released version 0.3, introducing five new text-to-speech models: Supertonic 3, MOSS-TTS-Local, MOSS-TTS-Nano, IndexTTS2, and Irodori-TTS. The update also adds GGUF support, which will be rolled out model by model.

• Supertonic 3 achieves over 200x real-time speed on CUDA (RTX 5090) and approximately 47 ms TTFT in streaming mode.
• A demo generated 10 hours of audio from The Adventures of Sherlock Holmes in about 3 minutes on an RTX 5090.
• Supertonic 3 is significantly faster than the official ONNX implementation because it avoids sending nodes back to the CPU.
• IndexTTS2 shows a 5.65x speedup over Python for long-form inputs (6000-char text + 2400-char emotion) on CUDA.
• Other new models range from similar speeds to more than 2x faster depending on the test case.

The release highlights substantial performance gains for C++/GGML inference compared to Python implementations, particularly on GPU hardware.