I developed Dictámelo, an open-source dictation app for macOS and Windows
It follows a small loop: hold a shortcut, speak, release, and the transcript appears at your cursor
The application is MIT licensed, including the Rust backend and Tauri interface
Choosing where transcription runs
The local catalog includes Whisper Tiny, Base, Small and Large v3, Parakeet v3 and Canary 180M Flash
Download a model once and local transcription works without an account, API key or internet connection, with no hosted time allowance
Cloud is a separate choice, either your own provider key or Dictámelo Cloud
That distinction also applies to cleanup: local speech recognition does not make every optional feature local, and AI text cleanup sends the transcript to a cloud cleaner only with consent
The desktop details
The recording shortcut is configurable, Escape cancels a recording, and the app keeps a local history so a failed paste does not mean losing the transcript
It can transcribe audio files and use a personal vocabulary for names and terms
The native code handles recording, model execution and platform integration, while Tauri hosts the settings and onboarding interface
For contributors, the useful starting points are src-tauri/src/pipeline.rs, the transcription adapters, and docs/LOCAL_MODELS.md
Trying it
- Download the installer for Apple Silicon macOS or Windows x64/ARM64
- Choose local models in setup
- Download a model and select Use model
- Hold Alt/Option + Shift + Space, speak, and release in a text field
Local use is free, personal cloud providers have their own rates and limits, and the hosted service has optional Free and Pro plans
I see it as an open-source alternative to Wispr Flow for people who want a simple dictation workflow and control over where speech is processed
Model speed and language coverage depend on the model and hardware, so I would especially value reports about first-run setup and everyday dictation in different languages
The app is MIT licensed, while third-party model weights retain their own licenses, documented in the repository
Top comments (1)
The local history detail — a failed paste not losing the transcript — is the kind of thing you only add after it bites you personally, so this reads like a tool built from usage rather than a feature list.
Since you asked for reports on everyday dictation in different languages: the failure mode I'd watch with the small Whisper tiers is hallucination on silence and background noise — the model confidently emitting a "Thank you." loop or an ad jingle when nobody spoke. Do you gate the audio through a VAD before it hits the model, or pass raw audio and handle it after? Related: what does Escape actually do — discard the buffer client-side, or transcribe and drop?
One question on the catalog: with Whisper Large v3, Parakeet v3 and Canary 180M Flash side by side, what's the perceived key-release-to-paste latency for a typical 10-second utterance on Apple Silicon? My experience is that number, more than accuracy, decides which model people leave enabled day to day.