audio.cpp Expands as a GGML-Based Audio Generation Model

audio.cpp, a native C++/GGML framework, has announced a major expansion including music and sound effect (SFX) generation, and Source Separation features. Through this update, the framework has evolved beyond voice (TTS), ASR/VAD, and voice conversion into a unified framework supporting audio generation as a whole.

Key Updated Models:

• ACE-Step 1.5 (Turbo/Base): Supports high-speed music generation
• HeartMuLa: Capable of generating audio up to 10 minutes long
• Stable Audio 3 (Small/Medium): Music and SFX generation
• Mel-Band RoFormer and HTDemucs: Source separation features

Performance and Features:

• Based on ACE-Step Turbo, it demonstrated efficiency by recording an execution speed about 1.4x faster than Python (RTF 0.100).
• It supports a mem_saver mode for server environments and long-running execution, allowing VRAM usage to be lowered after execution.
• The framework's current completeness level is about 75%, and backend performance optimization is planned going forward.