Local Whisper

**Quick judgment:** Solid, practical offline Whisper-based STT skill with good model flexibility and privacy properties. Well-suited as a local transcription backend, especially where network access or data privacy are concerns. **What it does well:** - Provides **fully offline speech-to-text** once models are downloaded, avoiding API keys, quotas, and network dependency. - Supports multiple model sizes (`tiny`, `base`, `small`, `turbo`, `large-v3`) so you can tune for speed vs. quality. - Offers useful CLI options: model selection, language override, timestamps, JSON output, and quiet mode. - Uses a reproducible Python environment (`uv` + `.venv`) and standard libraries (`openai-whisper`, `torch`). **Main risks / limitations:** - **CPU-only PyTorch wheel** (per the provided install command) may mean **slow transcription** on larger models, especially `turbo` and `large-v3` on long audio. - Requires a **local Python 3.12 environment** and some comfort with the command line; non-technical users may struggle without additional tooling. - Large models (809M–1.5GB) have **significant disk and RAM requirements**, which may be problematic on constrained devices. - No explicit mention of language coverage beyond Whisper’s defaults; quality may vary on low-resource languages. **Recommended scenarios:** - You need **reliable, high-quality STT** on a machine without stable internet or where cloud APIs are disallowed. - You want to **integrate STT into an automated pipeline**, using JSON + timestamps as machine-readable output. - You accept **higher latency** on CPU in return for **privacy and full local control**. **Less ideal for:** - Very resource-constrained machines or real-time needs where CPU-only Whisper will be too slow. - Users who require a polished GUI or one-click setup rather than a CLI-driven workflow.