audio.cpp 0.4 adds 5 speech models and reports up to 37% VRAM savings
Acceptable-Cycle4645 · reddit · 2026-07-24
audio.cpp 0.4 adds several new speech models and makes GGUF a first-class format across the project.
- New support includes Higgs Audio v3 TTS 4B, Fish Audio S2 Pro, Voxtral Realtime ASR, plus community ports OuteTTS and VieNeu-TTS-v3.
- The project now supports 35 model families, and every released family now has GGUF packages.
- Benchmarks on an RTX 5090 suggest Q8 GGUF can deliver real gains: Higgs Audio TTS runs about 8.8×–10.1× real time on warmed requests, Fish Audio S2 Pro about 3.1×–3.4×, and Voxtral ASR about 15.7× with streaming TTFT around 171 ms.
- Compared with 16-bit GGUF, Q8 can be up to 1.5× faster and cut peak VRAM by up to 37%, though quality remains model-dependent.
Related event: audio.cpp 0.4 Adds Voice Models and Boosts VRAM Efficiency(2 posts)→
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- SpaceX locks in $40B GPU financing deal with Apollo, the largest ever committed — TechCrunch AI · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27