Skip to content

audio.cpp Windows prebuilt binaries v0.5

Latest
Latest

Choose a tag to compare

View all tags
@0xShug0 0xShug0 released this 31 Jul 17:55
release-0.5
3178daf

These Windows prebuilt packages are for audio.cpp 0.4.2, built from commit 3178daf on main.

WebUI not included.

What's New

New Features

  • Added 9 model families, growing audio.cpp from 35 to 44 supported families:
    dramabox, confucius4_tts, rvc, bs_roformer, glm_tts, kroko_asr, parakeet_tdt, inflect_v2, and fun_asr_nano.
  • Added early HIP/ROCm backend support for AMD GPUs, including Linux HIP build support and Windows HIP distribution documentation.
  • Huge performance boost for selected models on Metal (>5x)
  • Added Qwen3 ASR streaming support.
  • Added live PCM server ingest through POST /v1/audio/transcriptions/live.
  • Added model spec v1 and model manager v2 groundwork: typed metadata, package links, dependency conditions, canonical package specs, and improved GGUF download defaults.
  • Added composite model build selection so builds can include only selected model families instead of always compiling the full model set.
  • Improved GGUF/package support:
    • Better standalone GGUF directory loading.
    • More model specs migrated to v1.
    • Package-backed downloads used more consistently by WebUI/model manager.
    • More tested GGUF precision paths, including Q4 for Voxtral Realtime.
  • Refreshed Docker documentation and examples for published images and GGUF usage.

Bug Fixes and Stability

  • Fixed build_windows.ps1 fallback CMake/Ninja discovery for standalone tools.
  • Fixed Qwen3 TTS decoder chunk graph reuse.
  • Fixed WebUI model download tuple mismatch.
  • Fixed WebUI Windows preset build discovery and improved GGUF/Metal handling.
  • Fixed standalone dependency model loading.
  • Hardened safetensors parsing with data_offsets validation.
  • Hardened file/header-controlled allocation sizes.
  • Hardened PocketTTS FlowLM step allocation against OOM.
  • Fixed PocketTTS clone audio path handling on Windows.
  • Improved IndexTTS number normalization and memory defaults.
  • Improved Supertonic multilingual text handling.
  • Improved Fish Audio server request reuse.
  • Improved Voxtral Realtime streaming behavior and prevented EOS from incorrectly ending a stream.
  • Preserved VibeVoice ASR turn text and aligned ASR chunk timestamp domains.
  • Changed streaming transcript output to emit deltas instead of repeatedly restating the transcript.

Packages

CPU packages are self-contained:

  • audiocpp-windows-cpu-fast.zip
  • audiocpp-windows-cpu-balance.zip
  • audiocpp-windows-cpu-portable.zip

CUDA packages are split into one shared runtime package plus one profile package:

  • audiocpp-windows-cuda-runtime.zip
  • audiocpp-windows-cuda-fast.zip
  • audiocpp-windows-cuda-balance.zip
  • audiocpp-windows-cuda-portable.zip

The CUDA runtime package contains the large CUDA DLLs shared by all CUDA profiles. Download it once, then pair it with whichever CUDA profile package you want to use. You don't need to download it if you already have it from previous releases.

Which Package Should I Download?

Use a CUDA package if you have a supported NVIDIA GPU. CUDA profile packages include both CUDA and CPU backend support, so they can run with either --backend cuda or --backend cpu once the CUDA runtime DLLs are available.

Use a CPU package if you do not have a supported NVIDIA GPU or only want CPU inference.

For most users:

  • NVIDIA GPU: download audiocpp-windows-cuda-runtime.zip and audiocpp-windows-cuda-balance.zip
  • CPU only: download audiocpp-windows-cpu-balance.zip

CPU Architecture Profiles

The package names use three CPU compatibility profiles:

Profile CPU build setting Best for Tradeoff
fast Native CPU optimization Newer high-end x86-64 CPUs that support the same instruction families used by the package Fastest CPU path, but least portable. If your CPU is older or missing newer SIMD instructions, use balance or portable.
balance AVX2 CPU kernels, native CPU optimization off Most modern x86-64 Windows PCs with AVX2 support Recommended default. More portable than fast, faster than portable, and avoids AVX512/AVX-VNNI requirements.
portable Baseline CPU kernels, native CPU optimization off, llamafile SGEMM off Older or unknown x86-64 Windows CPUs Broadest compatibility, but slowest. Avoids AVX, AVX2, AVX512, AVX-VNNI, and llamafile SGEMM.

The CUDA packages still include a CPU backend, so the same CPU profile applies when a CUDA package is run with --backend cpu.

CUDA Package Requirements

  • 64-bit Windows
  • NVIDIA GPU with compute capability 7.5 or newer
  • NVIDIA driver 580 or newer
  • Model files downloaded separately

The CUDA Toolkit and Visual Studio Build Tools are not required to run the CUDA packages. The required CUDA runtime DLLs are provided in audiocpp-windows-cuda-runtime.zip.

The CUDA packages are intended for RTX 20/30/40/50 series GPUs and similar NVIDIA datacenter GPUs. Older GPUs such as GTX 10-series Pascal cards or V100-class Volta cards are not covered by the CUDA 13 package; use a CPU package or build a separate package with an older CUDA Toolkit if those GPUs must be supported.

Quick Start

CPU users: unzip one CPU package, then run:

.\audiocpp_cli.exe --help

CUDA users: unzip audiocpp-windows-cuda-runtime.zip and one CUDA profile package into the same directory, then run:

.\audiocpp_cli.exe --help

CUDA example:

.\audiocpp_cli.exe --backend cuda --task tts --family <family> --model C:\path\to\model [options]

CPU example:

.\audiocpp_cli.exe --backend cpu --task tts --family <family> --model C:\path\to\model [options]

Server:

.\audiocpp_server.exe --config C:\path\to\server.json

Notes

  • Models are not bundled.
  • For CPU packages, keep the bundled DLL files next to the .exe files.
  • For CUDA packages, keep the CUDA runtime DLLs from audiocpp-windows-cuda-runtime.zip next to the .exe files, or put that runtime directory on PATH.
  • If CUDA startup fails, update the NVIDIA driver first.
Assets 9
Loading