Skip to content

Quick Start: GPU Acceleration

Get GPU acceleration working in 5 minutes.

Installation

Install with GPU Support

pip install voicepad-core[gpu]

This installs:

  • Base voicepad-core (~60MB)
  • CUDA 12 libraries (~500MB in your virtual environment)

Total: ~560MB (isolated to your venv, no system changes)

Verify GPU is Working

Run a test transcription:

voicepad record start

Check the output for Device: cuda (GPU working) or Device: cpu (using CPU fallback—see Troubleshooting).

How GPU Acceleration Works

Two Modes

CPU Mode (Default)

  • Works on all systems
  • ~8-10 seconds per minute of audio
  • Included with base package

GPU Mode (With [gpu] extra)

  • Requires NVIDIA GPU
  • ~2-3 seconds per minute of audio
  • 4-5x faster than CPU
  • Install: pip install voicepad-core[gpu]

Automatic Detection

At transcription time, voicepad checks for CUDA libraries:

  1. Attempts to import CUDA modules (nvidia.cublas.lib, nvidia.cudnn.lib)
  2. If available → uses GPU automatically
  3. If missing → falls back to CPU with helpful guidance

No manual configuration needed. The transcription_device: auto setting in your config does this automatically.

Virtual Environment Isolation

CUDA libraries live only in your virtual environment, not globally:

your-venv/Lib/site-packages/nvidia/cublas/  # CUDA here
your-venv/Lib/site-packages/nvidia/cudnn/   # CUDA here

Benefits:

  • ✅ Different projects can use different CUDA versions
  • ✅ No conflicts with system CUDA
  • ✅ Clean uninstall: just pip uninstall nvidia-cublas-cu12 nvidia-cudnn-cu12
  • ✅ Works even if system has CUDA 13 while venv uses CUDA 12

Common First Steps

Check if your GPU is detected:

voicepad config system

Look for:

GPU:
   NVIDIA Driver: [OK] Detected
   CUDA Devices: 1 device(s) available
   faster-whisper GPU: [OK] Compatible

Get a model recommendation for your system:

voicepad config recommend

See Advanced Setup for hardware requirements and detailed configuration.