whisper.cpp-release

Prebuilt whisper.cpp binaries for Linux, Windows, and macOS — built and released via GitHub Actions.

Each release includes whisper-cli and whisper-server.

Platform support

Platform	Arch	GPU backend
Linux	x86_64	Vulkan (release), SYCL (local build)
Windows	x86_64	Vulkan
macOS (Apple Silicon)	arm64	CoreML
macOS (Intel)	x86_64	CPU only

Download

Get the latest binaries from GitHub Releases.

Python CLI (whisper_tool)

A Python helper that downloads default models, locates or fetches prebuilt binaries, and exposes common operations via CLI.

Requirements: Python 3.10+, stdlib only (no pip install needed).

Run the launcher script from the repo — works from any working directory:

cd whisper.cpp-release

# One-shot setup: binaries (from local build/release or GitHub Releases) + default models
./bin/whisper-tool setup

# Check resolved paths
./bin/whisper-tool status

# Transcribe (VAD enabled by default)
./bin/whisper-tool transcribe audio.wav

# Start HTTP server
./bin/whisper-tool serve --host 0.0.0.0 --port 8080

Install to PATH (optional):

ln -s "$(pwd)/bin/whisper-tool" ~/.local/bin/whisper-tool
whisper-tool status

Equivalent without the launcher: python3 -m whisper_tool ... (must run from repo root, or set PYTHONPATH).

Another optional install via pip: pip install -e . / pipx install -e . also provides a whisper-tool command.

Default models

Type	Model	File
Whisper	small.en	ggml-small.en.bin
VAD	silero-v6.2.0	ggml-silero-v6.2.0.bin

Default data locations:

Platform	Root directory
Linux / macOS / WSL	~/.local/share/whisper-cpp-release/
Windows	%LOCALAPPDATA%\whisper-cpp-release\

Under the root: models/, bin/ (Linux/macOS/Windows), bin-win/ (WSL Windows .exe), downloads/. Config is config.json in the root on Windows, or ~/.config/whisper-cpp-release/config.json on Linux/macOS/WSL.

WSL: use Windows binaries for Vulkan GPU

On WSL, the tool automatically prefers Windows whisper-cli.exe / whisper-server.exe so inference can use Windows Vulkan drivers. setup downloads the windows-x64 release into bin-win/. Model and audio paths are converted via wslpath when calling .exe.

./bin/whisper-tool setup      # downloads windows-x64 binaries on WSL
./bin/whisper-tool status     # shows wsl: true, windows_binary: true
./bin/whisper-tool transcribe audio.wav

Force Linux binaries (CPU only): WHISPER_TOOL_LINUX_BINARIES=1 ./bin/whisper-tool setup

Intel GPU: SYCL backend (local build)

SYCL is not built or published in GitHub Actions releases (the CI compile is too memory-heavy). Build locally on Intel CPU/GPU systems with oneAPI installed:

source /opt/intel/oneapi/setvars.sh   # or use Nix: nix shell nixpkgs#intel-oneapi-toolkit ...
GGML_SYCL=1 make build
GGML_SYCL=1 make release            # release/whisper-<version>-<arch>-sycl.tar.gz

Then point whisper-tool at the local build (copy into bin/, or use a local release/whisper-*-sycl/ directory). Runtime requires Intel oneAPI + Level Zero on the target machine.

./bin/whisper-tool status   # discovers local release/whisper-*-sycl/ if present

Commands

./bin/whisper-tool setup                              # binaries + default models
./bin/whisper-tool download binary                    # fetch release binaries only
./bin/whisper-tool download whisper [MODEL]           # e.g. base.en, small.en
./bin/whisper-tool download vad [MODEL]               # e.g. silero-v6.2.0
./bin/whisper-tool transcribe FILE [FILE...] [--no-vad] [-l LANG] [--output-json]
./bin/whisper-tool transcribe audio.wav -- -pp          # pass extra args to whisper-cli after --
./bin/whisper-tool serve [--host HOST] [--port PORT] [--no-vad]
./bin/whisper-tool status
./bin/whisper-tool clean [--dry-run] [--models] [--binaries] [--config]

clean removes downloaded user data from the platform-specific directories above. Without flags it removes everything; use --dry-run to preview.

Binary discovery order:

• WSL (default): bin-win/ → whisper-cli.exe on PATH → release/whisper-*-windows-x64/ → fallback to Linux bin/
• Linux/macOS: config bin_dir → PATH → whisper.cpp/build/bin/ → release/whisper-*/ → downloaded bin dir

Build locally

git clone --recurse-submodules https://github.com/jiang1997/whisper.cpp-release.git
cd whisper.cpp-release
make build

Override GPU support:

GGML_VULKAN=0 make build      # CPU-only on Linux/Windows
GGML_SYCL=1 make build        # Intel SYCL on Linux/Windows (oneAPI required)
WHISPER_COREML=0 make build   # CPU-only on macOS ARM

Release

make release

Produces release/whisper-<version>-<arch>.tar.gz containing the CLI, server, and shared libraries.

Acknowledgements

The cross-platform build strategy (platform-specific CMake flags, GPU backend selection, CI build matrix) is based on Buzz, an offline transcription desktop app that bundles whisper.cpp binaries across Linux, Windows, and macOS.

License

The build scripts in this repository are MIT-licensed. whisper.cpp itself is MIT-licensed.