Skip to content

armgit / Handy

Public Imported from GitHub
0

A free, open source, and extensible speech-to-text application that works completely offline.

2.9 KB
Raw Blame

Transcribing Files

Handy is now able to transcribe files! You can do this from the history tab and click the "Transcribe File" button, or just drag and drop any file directly onto Handy's window

History tab with the Transcribe File button

Preferred Languages and New Model Picker

You can now tell Handy which languages you speak, and it will use that to recommend models for you. You can change them any time in General → Preferred Languages.

The model picker has been rebuilt around this. Every model now gets an accuracy grade (A through F) for each of your languages. Handy picks a Recommended model, plus a Faster and a More accurate option, so you don't have to dig through every model to find a good one.

Models page with accuracy grades for English and Chinese

Auto Language Detection

Auto now works with all transcription models. Some models can't tell which language you're speaking on their own, so Handy uses a small language detector to figure it out. If you've selected a model which can't tell your language, you can download the detector by picking Auto in the language dropdown.

There is a new experimental setting to restrict auto to only your preferred languages. If you want that, go check it out!

ONNX Models No Longer Supported

For people still using ONNX based models, support for them has been removed. There is some UI to help guide you to upgrading to the new type of model, and it should work just as well on your system, if not even better, than before. If you have any issues, please feel free to contact [email protected] or leave a Github Issue

Other Changes

  • More stable: transcription now runs in its own separate process, so if the model or your GPU crashes or hangs, Handy itself keeps running and recovers. If the GPU was the problem, it will retry on the CPU.
  • Better errors: they will show up in the overlay immediately as well as a tray in the app UI.
  • Language picking in the menu bar: as it sounds, you can change your model language from the menu bar.
  • Better long audio handling: By default, long audio, will be chunked, and this, should help long audio transcription performance as well as prevent some models from crashing. Can be turned off in Advanced → Long Audio Chunking.
  • New onboarding: it should smooth things out quite a lot and make it easier for people to get started with local models.
  • Custom words work with more models: Qwen3 ASR, Granite and others now get your custom words directly, not just Whisper.
  • AirPods Mode (macOS, experimental): fixes recordings coming back silent when your iPhone is playing audio through your AirPods.
  • Lots of smaller fixes, including better keyboard and screen reader support across settings.