Skip to content

GPU Acceleration

Local transcription with Whisper models can use your GPU to significantly reduce transcription time. OSTT supports Metal on macOS, CUDA on NVIDIA Linux hardware, and Vulkan on AMD and Intel Linux hardware.

GPU acceleration has no effect on cloud providers — it only applies to local models.

Backends

BackendHardwarePlatform
MetalApple Silicon and Intel MacmacOS
CUDANVIDIA GPUsLinux x86_64
VulkanAMD and Intel GPUsLinux x86_64

macOS — Metal

Metal is enabled automatically on every macOS build. No setup is needed. When a local model runs, OSTT uses the GPU without any configuration.

Verify it is active in the logs:

bash
ostt logs

Look for:

local transcription: Metal GPU acceleration enabled

Linux — NVIDIA (CUDA)

The CUDA build is available for Linux x86_64 and requires:

  • An NVIDIA GPU
  • NVIDIA driver (libcuda.so)
  • cuBLAS runtime (libcublas.so)

The install script detects usable NVIDIA CUDA support automatically and installs the CUDA build:

bash
curl -fsSL https://ostt.ai/install | bash

To opt out of GPU detection and install the CPU build:

bash
curl -fsSL https://ostt.ai/install | bash -s -- --no-gpu

Verify CUDA is active:

bash
ostt logs

Look for:

local transcription: CUDA GPU acceleration enabled

Linux — AMD and Intel (Vulkan)

The Vulkan build is available for Linux x86_64 and works on AMD and Intel GPUs via the Mesa driver stack. libvulkan.so.1 ships with Mesa on most graphical Linux installs, so no additional runtime dependencies are typically needed.

The install script detects AMD and Intel GPUs via lspci and installs the Vulkan build automatically:

bash
curl -fsSL https://ostt.ai/install | bash

To opt out of GPU detection and install the CPU build:

bash
curl -fsSL https://ostt.ai/install | bash -s -- --no-gpu

Verify Vulkan is active:

bash
ostt logs

Look for:

local transcription: Vulkan GPU acceleration enabled

Manual Package Selection

If you install from .deb, .rpm, or direct release archives instead of the installer, choose the artifact that matches your hardware:

HardwareDebian/UbuntuFedora/RHEL/openSUSEArchive
CPU or Linux ARM64ostt_latest_amd64.deb or ostt_latest_arm64.debostt-latest.x86_64.rpm or ostt-latest.aarch64.rpmostt-<target>.tar.gz
NVIDIA GPU with CUDA runtime librariesostt-cuda_latest_amd64.debostt-cuda-latest.x86_64.rpmostt-x86_64-unknown-linux-gnu-cuda.tar.gz
AMD/Intel GPU with Vulkan runtime libraryostt-vulkan_latest_amd64.debostt-vulkan-latest.x86_64.rpmostt-x86_64-unknown-linux-gnu-vulkan.tar.gz
macOSHomebrew or macOS archiven/aostt-x86_64-apple-darwin.tar.gz or ostt-aarch64-apple-darwin.tar.gz

Run uname -m to check your CPU architecture. CUDA and Vulkan artifacts are currently Linux x86_64 only.

Detection Priority

On Linux, if both NVIDIA and AMD/Intel GPUs are present, the install script selects the CUDA build. CUDA is faster for NVIDIA hardware than Vulkan.

CPU Fallback

The CPU build works on all hardware. If you install the GPU build and the required runtime libraries are not present, the binary will fail to start. In that case, reinstall with --no-gpu or install the required driver packages.