[audio.cpp] VibeVoice 1.5B released — 90-min podcast in 22.95 min, 4.08x real-time, 2.86x faster than Python without quantization. Native C++/ggml

I’m the author of audio.cpp, a C++/ggml runtime for local audio models.

I just added VibeVoice 1.5B support and wanted to share the benchmark because long-form multi-speaker TTS is a good stress test for local inference runtimes.

Result on RTX 5090:

VibeVoice 1.5B Audio length: 5615.73s / 93.60 min Wall time: 1376.84s / 22.95 min RTF: 0.245 Speed: 4.08x faster than real time Python baseline: 92.66 min audio in 65.70 min Speedup vs baseline: 2.86x Quantization: none Diffusion steps: 10

The main point is not just avoiding Python setup pain, though that is part of it. The goal is to make audio models practical in a native local runtime: reusable sessions, server-like usage, long-form generation, stable memory behavior, and CUDA-focused (CPU and Metal later) optimization.

VibeVoice is a useful milestone because it is not just short-sentence TTS. It is designed for long-form, multi-speaker dialogue such as podcasts, character chats, and narration, where runtime behavior matters a lot.

Current framework progress:

Released model families: 16 / 28 [███████████░░░░░░░░░] 57%

The other model families are already running end-to-end internally, but I’m releasing them gradually after testing and cleanup.

The repo is https://github.com/0xShug0/audio.cpp

I’d be interested in feedback from people testing VibeVoice on other GPUs or CPUs, especially long prompts, multi-speaker formatting, VRAM behavior, and performance numbers.