Presspeech vs Handy
Both are free, MIT-licensed, local dictation apps: trigger a recording, speak, and put the transcript into the active app. Handy prioritises cross-platform reach, model choice, and extensibility. Presspeech uses separate native pipelines with fewer choices and a measured macOS latency target.
| Question | Presspeech | Handy |
|---|---|---|
| Primary workflow | Hold-to-talk or toggle, then normally paste after confirming the same window is still focused, with manual clipboard recovery when needed. On the published Mac benchmark, model transcription takes 92–152 ms p50 across the tested clips; clipboard and paste work are not timed. | Hold-to-talk or toggle, then paste the transcript into the active text field. |
| Platforms | Released native app for Apple Silicon and macOS 14+; separate Windows preview for x64 PCs. | One Tauri application for Apple Silicon and Intel Macs, x64 and Arm64 Windows, and x64 and Arm64 Linux. |
| Speech engines | Fixed per platform: multilingual Parakeet TDT v3 through FluidAudio/CoreML on macOS; multilingual Parakeet TDT v3 on NVIDIA CUDA or English-only Whisper base.en by default on CPU-only Windows PCs. | User-selectable local Whisper models and Parakeet models. Its documentation describes GPU acceleration for Whisper and CPU operation for Parakeet. |
| Paste and clipboard recovery | Checks that the captured destination is still the focused window before automatic paste; otherwise it keeps the transcript available for explicit manual paste. On macOS 0.3.8, optional preservation of the previous clipboard requires a separate Restore action and confirmation rather than restoring on a timer. | Handy's documented standard clipboard method restores the previous clipboard after a fixed delay. Its troubleshooting guide warns that a slow target can read the old content instead; macOS and Windows also offer a Reliable Paste beta that waits for a clipboard-read notification, plus adjustable delays for the standard method. On Linux, its README notes the recording overlay is off by default because some compositors treat it as active; enabling it may steal focus and make clipboard pasting fail or go to the wrong window. |
| Privacy posture | No cloud transcription, Presspeech-authored analytics, account, or crash reporter. Speech-library network behavior varies by published platform build: Windows 0.1.12 leaves Hugging Face libraries' default usage telemetry enabled during model downloads. See the version-specific network-call inventory. Dictation audio is held in memory and discarded. | Core transcription happens on the computer without uploading audio. Handy's privacy policy says update checks are enabled by default and model downloads may fall back from Hugging Face to its Cloudflare-delivered model host; optional post-processing can call a configured local or cloud provider. |
| Text handling | Deterministic dictionary corrections, shortcuts, and filler removal on both platforms; macOS also has opt-in spoken formatting. No AI rewriting. | Core local transcription plus model and language selection, a dictionary, filler-word removal, and optional post-processing controls. |
| Transcript and audio retention | On macOS, optional recent transcripts stay in memory and clear on quit. Windows has no in-app transcript history. Completed audio is discarded on both platforms; the system clipboard and third-party clipboard managers have separate retention behavior. | Handy's privacy policy says official releases keep a limited number of recent local history entries and their audio recordings by default. Saved entries may remain until deleted; users can change retention or set the history limit to zero. |
| Implementation and scope | Native Swift/AppKit macOS app with a separate Python Windows implementation. The macOS release zip is 8.4 MB before its model download. | Extensible Tauri, React, and Rust application shared across platforms. Current release packages and models vary by operating system and architecture. |
| Price and license | Free and MIT licensed. | Free, MIT licensed, no account. |
Choose Handy if Linux or Intel Mac support, one cross-platform application, model switching, or a configurable local transcript and audio history matters most. Choose Presspeech if you want the smaller native macOS surface and its published model-transcription benchmark, or its intentionally fixed engine and no persistent transcript archive. On Windows, test both on the actual hardware: their packaging and acceleration paths differ substantially.
Sources: Presspeech source and platform documentation, Presspeech benchmarks, Handy's Linux notes and clipboard troubleshooting guidance (checked 22 October 2026), its privacy policy (effective 18 September; checked 24 September 2026), and the Handy v0.9.8 release and release assets (checked 22 October 2026). These are the projects' published claims, not independent privacy or performance tests. No cross-product latency or accuracy result is claimed because neither project publishes a controlled head-to-head benchmark.