audio.cpp v0.3: Generate 10 hrs audio in 3 min on RTX 5090
Supertonic 3 hits 200x real-time; 10 hours of audio in 3 minutes on RTX 5090
audio.cpp 0.3 introduces five new text-to-speech models: Supertonic 3, MOSS-TTS-Local, MOSS-TTS-Nano, IndexTTS2, and Irodori-TTS. The standout is Supertonic 3, which achieves over 200x real-time speed on an RTX 5090 (and 6x on CPU) with a time-to-first-token (TTFT) of just 47ms in CUDA streaming mode. In a demo using Sherlock Holmes as input, it generated roughly 10 hours of audio in about 3 minutes on the high-end GPU. The developer reverse-engineered the original ONNX-based inference to rebuild it with safetensors weights, resulting in dramatically faster performance because the C++ version keeps all operations on the GPU instead of shunting nodes back to the CPU.
Other models in the release also show significant gains. IndexTTS2, when tested with a long input (6000-character text plus 2400-character emotion control), ran 5.65x faster than the Python reference implementation. The project now includes early GGUF support (being rolled out model by model) and remains fully open source with a focus on expanding model coverage, streaming, and UI/server layers. Contributions are welcome as the developer continues to push local, high-speed TTS generation.
- Supertonic 3 achieves 200x+ real-time on RTX 5090, generating 10 hours of audio in 3 minutes.
- C++/GGML implementation outperforms Python ONNX by up to 5.65x for IndexTTS2 on longform tasks.
- Release adds five new TTS models and early GGUF support, with more improvements planned.
Why It Matters
audio.cpp makes high-speed, local TTS generation accessible, enabling near-instant audio creation for developers and creators.