audio.cpp Update Brings Insanely Fast Audio Generation
Acceptable-Cycle4645 · reddit · 2026-07-15
audio.cpp releases version 0.3, adding 5 new models: Supertonic 3, MOSS-TTS-Local, MOSS-TTS-Nano, IndexTTS2, and Irodori-TTS.
Key Highlights
- Supertonic 3 achieves 200x+ real-time speed on CUDA and 6x+ on CPU, with a TTFT of 47ms in CUDA streaming mode
- A demo using The Adventures of Sherlock Holmes generated 10 hours of audio in roughly 3 minutes on an RTX 5090
- The author reverse-engineered the official ONNX inference path and rewrote it using safetensors weights
- The CUDA version is significantly faster than the Python implementation, while CPU performance is on par or slightly faster
Other Notes
- Fallback to CPU for slow nodes is a major reason for sub-optimal GPU utilization in the ONNX version
- Speed improvements for other models in v0.3 range from near-baseline to 2x+
- In long-text tests for IndexTTS2, the C++ version is 5.65x faster than Python
- GGUF support has been added and will be rolled out progressively per model
Project repo: github.com/0xShug0/audio.cpp
Related event: audio.cpp Update Boosts Local Audio Generation Speed by 200x(2 posts)→
More from coding & agent
- Spring AI 2.1.0-RC1 adds tool-call confirmation, pausing agent loops before real-data access — therealdanvega · 2026-10-26
- Morgan Stanley and US banks block third-party AI agents over credentials — thione · 2026-10-26
- METR red-teams OpenAI's agent monitoring, finds bypasses and gaps — thione · 2026-10-26
- OpenAI: coding agents weakened safety monitoring, one lied about it — thione · 2026-10-26
- ConvFill: A 0.6B Talker Buys Time While a Large Reasoner Whispers the Real Answer — andimarafioti · 2026-10-26
- 250 Builders Hack With Codex at Berlin Community Hackathon Ahead of OpenAI DevDay — paw_lean · 2026-10-26