audio.cpp 0.4 delivers 10x real-time TTS with new GGUF support

audio.cpp, the open-source C++/GGML inference engine for audio models, has released version 0.4 with major updates. The headline additions are Higgs Audio v3 TTS (4B parameters) and Fish Audio S2 Pro, both now supported in GGUF format. The project also introduces Voxtral Realtime ASR and two community models — OuteTTS TTS and VieNeu-TTS-v3. With this release, audio.cpp covers 35 model families, all of which now support GGUF loading. The Q8 quantized format is starting to show real-world benefits: on an RTX 5090, Higgs Audio TTS achieves 8.8x–10.1x real-time speed for warmed requests and 8.5x for longform (6000+ characters). Fish Audio S2 Pro runs 3.1x–3.4x for warmed requests and 3.3x for longform. Voxtral ASR offline runs 15.7x real-time, with streaming time-to-first-token around 171 ms.

GGUF loading becomes a first-class feature, with ready-to-use packages available. Q8 quantization offers up to 1.5x speed improvement and up to 37% reduction in peak VRAM, depending on the model and route. However, the developer notes that quality is model-specific, and the support matrix is kept transparent. Some tricks (e.g., adjusting chunk size or reference audio length for Qwen3-TTS) can further boost performance up to 2x. audio.cpp also introduces a dedicated community models area for ports that are useful but still maturing, lowering the barrier for contributions. The project remains focused on efficient local inference for text-to-speech and speech recognition.