Forge

Forge / Voice & Audio

whisper.cpp

Open source

Local speech-to-text. The transcription plan that still works when Gemini is blocked and you cannot upload audio to a US API.

No cardMM-friendlySelf-hostable
Official siteGitHub

Free tier

MIT. Unlimited once the ggml model is on disk.

Limits

CPU-only is slow on large files. Burmese WER varies. First model download is heavy.

Best for

  • Interview tapes
  • Meeting notes
  • Offline captions

How to use it well

  1. Build from the repo or grab a release binary. Pull `base` or `small` first, not `large`.
  2. Convert audio to 16 kHz wav. Then `./whisper-cli -m models/ggml-small.bin -f talk.wav`.
  3. For Burmese, test `small` vs `medium`. If word error is high, transcribe in chunks.

Myanmar notes

Download models on Wi-Fi. `small` fits a Yangon laptop. Never upload sensitive venue audio to a cloud STT.