Advanced Setup

Hardware requirements, configuration, and technical details.

Hardware Requirements

GPU Support

Minimum: GTX 10-series or equivalent (Compute Capability 6.1+)

Software Requirements

• Windows or Linux (macOS Intel with eGPU may work)
• NVIDIA GPU Driver (any recent version)
• Python 3.13+
• Virtual environment recommended

Configuration

Check Current Settings

voicepad config show

Look for transcription settings:

transcription_device: auto        # auto, cuda, or cpu
transcription_compute_type: auto  # auto, float16, int8, or float32
transcription_model: medium

Force CPU Mode

To use CPU instead of GPU (if GPU is available):

# voicepad.yaml
transcription_device: cpu
transcription_compute_type: int8

Optimize for Your GPU

Recommendation: Run voicepad config recommend to get system-specific suggestions.

For manual tuning:

• Limited VRAM (<2GB): Use small model with int8 compute type
• Moderate VRAM (2-4GB): Use medium model with auto or float16
• Plenty VRAM (>4GB): Use large-v3 model with float16

Technical Details

What Gets Installed

Base package (voicepad-core):

• faster-whisper (transcription engine)
• CPU support included
• ~60MB total

GPU extra ([gpu]):

• nvidia-cublas-cu12>=12.0.0 (~200MB)
• nvidia-cudnn-cu12>=9.0.0 (~300MB)
• Total: ~560MB (added to your venv)

Virtual Environment Isolation

CUDA libraries are installed only in your virtual environment, not system-wide:

your-venv/
└── Lib/site-packages/nvidia/
    ├── cublas/bin/cublas64_12.dll     # Windows
    └── cudnn/bin/cudnn64_9.dll        # Windows

How detection works:

At transcription time, voicepad attempts to import:

import nvidia.cublas.lib
import nvidia.cudnn.lib

If both imports succeed → GPU mode enabled. If either fails → fallback to CPU.

Why Virtual Environment Isolation?

✅ No global system changes ✅ Different projects can use different CUDA versions without conflict ✅ Works even if system CUDA version differs from venv CUDA version ✅ Clean uninstall: pip uninstall nvidia-cublas-cu12 nvidia-cudnn-cu12 ✅ Reproducible builds across different systems

CUDA Compute Type Reference

Back to: Quick Start | Troubleshooting