Core Concepts
Transcription & Models
Local and cloud transcription backends, refinement models, and how to configure them.
VoiceWiz's pipeline is model-agnostic by design: transcription and refinement are configured independently, and both support local or cloud options. The Models screen in the Control Panel is where you choose.
Local transcription
Whisper-class models (whisper.cpp)
Whisper models run entirely on your machine, with CPU inference and optional CUDA GPU acceleration on NVIDIA hardware. Model sizes:
| Model | Size | Speed / Accuracy |
|---|---|---|
| Tiny | ~75 MB | Fastest / basic — quick notes |
| Base | ~142 MB | Fast / good — recommended balance |
| Small | ~466 MB | Moderate / better |
| Medium | ~1.5 GB | Slow / high accuracy |
| Turbo | ~1.6 GB | Fast / high accuracy — best tradeoff |
| Large | ~3 GB | Slowest / best accuracy |
NVIDIA Parakeet (sherpa-onnx)
Parakeet models use INT8-quantized ONNX inference on CPU:
| Model | Size | Languages |
|---|---|---|
| parakeet-tdt-0.6b-v3 | ~680 MB | 25 languages |
| parakeet-unified-en-0.6b | ~631 MB | English only |
Cloud transcription (bring your own key)
Cloud backends are optional. When configured, audio is sent to the provider you choose using your own API key:
- OpenAI — Whisper API transcription endpoints.
- Groq — fast Whisper-compatible inference.
- Mistral — supported transcription API.
- Custom endpoint — any compatible API you configure.
Refinement models
Local models
- llama.cpp — run GGUF models locally; VoiceWiz manages the server process for you.
- Ollama — use models managed by your local Ollama installation.
- LAN endpoint — any OpenAI-compatible endpoint on your network.
Cloud providers (bring your own key)
- OpenAI and OpenAI-compatible custom endpoints
- Anthropic (Claude)
- Google Gemini
- Groq
- DeepSeek
- OpenRouter
Cloud provider API keys are stored encrypted on your machine and hydrated at launch — they never leave your local storage except to call the provider you configured.
Choosing between local and cloud
- Local models need a one-time download and then work without a network connection.
- Cloud transcription can offload processing from slower machines — audio is sent to your configured provider.
- You can switch backends at any time from the Control Panel.
Mixing and matching
The two stages are independent. A common setup: local whisper for transcription (fast, private, offline-capable) plus a cloud provider for refinement (higher quality rewriting). The reverse also works.
Hardware notes
- Whisper-class models run on CPU; NVIDIA GPUs with 4 GB+ VRAM get CUDA acceleration automatically.
- Parakeet models are optimized for CPU inference.
- Local refinement models benefit from RAM and GPU headroom — start with a small model and scale up.