audio.cpp 0.4 Released with Higgs Audio v3, Fish Audio S2 Pro, and GGUF Support
The release of audio.cpp version 0.4 introduces C++/GGML implementations for Higgs Audio v3 TTS 4B, Fish Audio S2 Pro, and Voxtral Realtime ASR, alongside full GGUF format support across all 35 supported model families. It also introduces Q8 quantization, which delivers significant speedups and memory savings. This update significantly advances local audio AI by enabling high-performance, low-latency text-to-speech and speech recognition on consumer hardware without relying on heavy Python frameworks. The integration of state-of-the-art models like Higgs Audio v3 and Fish Audio S2 Pro makes real-time, expressive voice generation highly accessible. Tested on an RTX 5090, Q8 quantization reduces peak VRAM usage by up to 37% and increases speed by up to 1.5x compared to 16-bit GGUF. Higgs Audio TTS achieved up to 10.1x real-time performance, while Fish Audio S2 Pro ran up to 3.4x faster than real-time.
## BACKGROUND
GGML is a tensor library for machine learning designed for high performance on commodity hardware, while GGUF is its successor binary file format optimized for fast loading and saving of models. Higgs Audio v3 and Fish Audio S2 Pro are state-of-the-art open-source text-to-speech (TTS) models known for their expressiveness, multilingual capabilities, and low latency.