1. Whisper the model, Whisper the API, Whisper the app
OpenAI’s Whisper paper (December 2022) describes a family of speech recognition models trained on 680,000 hours of multilingual audio, released together with inference code so anyone can run them. Because the weights are open, Whisper turns up in many app names and feature lists. The model card lists 99 languages, but OpenAI says the models show strong results in about 10 of them and perform unevenly across languages, accents and dialects.
The same name also covers OpenAI’s hosted API. whisper-1 still exists for file transcription with a 25 MB upload limit, while OpenAI’s current documentation recommends its newer gpt-transcribe model and notes that whisper-1 does not support streamed output. An app that sends your recording to that API is cloud dictation, even if it says Whisper on the download page.
Before installing anything, check two things: whether the app runs the model on your computer, and which engine it uses. The first decides where your audio goes; the second decides how fast a given Whisper size runs on your hardware.
- Local Whisper: the model file sits on your disk and audio stays on the computer.
- API Whisper: each recording is uploaded to OpenAI for transcription.
- Hybrid apps: local models by default, with cloud engines you can switch on.
2. whisper.cpp, faster-whisper, MLX and WhisperKit
whisper.cpp is a plain C/C++ implementation with no dependencies. It treats Apple silicon as a first-class target through NEON, Accelerate, Metal and Core ML, and also supports x86 AVX, NVIDIA and AMD GPUs, Vulkan and CPU-only inference on macOS, Windows and Linux. It adds integer quantization and a voice activity detection option that sends only detected speech to Whisper. VoiceInk, Speech Note and OpenWhispr all credit it.
faster-whisper reimplements Whisper on the CTranslate2 inference engine and says it is up to 4 times faster than OpenAI’s reference code at the same accuracy, using less memory, with 8-bit quantization on CPU and GPU. In its own CPU benchmark (13 minutes of audio, the small model, an Intel Core i7-12700K with 8 threads), OpenAI’s code took 6 min 58 s, whisper.cpp 2 min 5 s, and faster-whisper with int8 1 min 42 s. Its optional Silero VAD filter, once enabled, removes silences longer than 2 seconds by default. It is a Python library, and Speech Note offers it as an engine on Linux.
Apple’s MLX is an array framework for machine learning on Apple silicon, and its mlx-whisper package runs Whisper checkpoints from 39 million to 1.5 billion parameters from the command line or Python. WhisperKit, from Argmax, is a Swift framework for on-device Whisper on Apple platforms. None of these engines is a dictation app on its own: they transcribe audio you hand them, and the app supplies the hotkey, the microphone capture, and the step that pastes or types text into the focused window.
3. Choosing a Whisper size for push-to-talk
OpenAI publishes approximate memory needs and speeds relative to large, measured transcribing English on an A100 GPU, and warns that real-world speed varies widely with language, speaking rate and hardware. The English-only .en versions of tiny, base, small and medium tend to be more accurate for English, with the gap shrinking at small and medium. Turbo is large-v3 with a 4-layer decoder, faster with minimal accuracy loss, and it is the default in OpenAI’s own command-line tool, though it is not trained for translation.
Quantized builds shrink those figures on a laptop. whisper.cpp lists small at 466 MiB on disk and about 852 MB of memory, medium at 1.5 GiB and about 2.1 GB, and large at 2.9 GiB and about 3.9 GB. On the Hugging Face Open ASR Leaderboard, read October 3, 2026, large-v3 averages 6.5% English WER at an RTFx of 470 and turbo 7.03% at 795, so turbo gives up about half a point of WER for roughly 70% more throughput.
Smaller models can still do well on clear speech. IraVoice’s July 2026 benchmark measured whisper.cpp small.en at 4.28% WER on a 40-speaker clean read-speech snapshot on an Apple M5 Max; it did not cover noisy, accented or spontaneous speech, and it did not test Parakeet. For live dictation, pick the largest model that still returns text without a pause you notice, then move down a size if it lags.
- tiny: 39M parameters, about 1 GB VRAM, about 10 times large’s speed.
- base: 74M, about 1 GB, about 7 times.
- small: 244M, about 2 GB, about 4 times.
- medium: 769M, about 5 GB, about 2 times.
- large: 1,550M, about 10 GB, the baseline.
- turbo: 809M, about 6 GB, about 8 times.
4. Local Whisper dictation apps on macOS
superwhisper runs on Mac and Windows, with iOS and Android apps, and says it works offline; its free plan includes unlimited use of Whisper models. Its FAQ says offline models run really well only on Apple silicon Macs and that Intel Macs work best with its cloud models. It also offers cloud models and AI modes that rewrite your text, so check which mode is active before dictating anything sensitive.
MacWhisper’s product page leads with transcribing audio files, and it also lists system-wide dictation with Whisper as a replacement for Apple’s own dictation, says all transcription is done on the device, and adds Parakeet v2 and v3 on M-series Macs. VoiceInk is a native app for macOS 15 or later whose source is published under GPL v3; it credits whisper.cpp for Whisper and FluidAudio for Parakeet and describes its processing as fully offline.
Two cross-platform options also run on the Mac. Spokenly runs local Whisper and Parakeet models for free, supports cloud engines through your own API keys, and has a local-only mode that blocks every network request. Handy is free and MIT-licensed: press a shortcut, speak, release, and it transcribes with Whisper Small, Medium, Turbo or Large, using GPU acceleration when available, or with Parakeet V3.
5. Whisper dictation on Windows and Linux desktops
On Windows, SpeechPulse uses Whisper voice recognition, supports offline recognition, types into any text input area, transcribes 99 languages and supports push-to-talk with custom hotkeys. superwhisper, Handy, Spokenly and OpenWhispr also ship Windows builds. OpenWhispr is MIT-licensed, runs local Whisper, NVIDIA Parakeet and other models or cloud providers, and advertises GPU-accelerated local Whisper on Metal, CUDA and Vulkan, which covers PCs with AMD or Intel graphics. Windows’ built-in voice typing (Windows key + H) needs no setup but uses online recognition powered by Azure Speech services, so it needs an internet connection.
On Linux, Speech Note (net.mkiol.SpeechNote on Flathub) processes everything offline and offers whisper.cpp and Faster Whisper among its engines, with optional GPU add-ons. Its Insert into active window mode makes it a voice typing tool: it works out of the box on X11, while Wayland needs the ydotool daemon running and, under Flatpak, permission to reach its socket. Handy, Spokenly and OpenWhispr have Linux builds too; Handy’s README says Wayland text input needs wtype or dotool.
IraVoice for Linux is a beta, shipped as an AppImage for any 64-bit distribution from 2022 on or as a .deb or .rpm, and it offers the same optional Whisper models as the Mac and Windows apps. On Linux the default shortcut is Ctrl+Space, and the text is typed through a helper such as wtype, ydotool or xdotool depending on your desktop.
- Windows: SpeechPulse, superwhisper, Handy, Spokenly, OpenWhispr.
- Linux: Speech Note, Handy, Spokenly, OpenWhispr.
- Wayland: expect to install a typing helper whichever app you choose.
6. When Parakeet is the better default than Whisper
Whisper is the more general model: 99 listed languages, a language setting you can fix, and translation into English on the non-turbo models. For the most common dictation case, English or a major European language spoken into a computer microphone, NVIDIA’s Parakeet TDT 0.6B v3 is the stronger default on published numbers: 5.66% average English WER at an RTFx of 6,076 on the Open ASR Leaderboard, against 7.03% at 795 for Whisper turbo, with 600 million parameters instead of 809 million. Our Parakeet vs Whisper guide covers the full comparison, including the languages where Whisper large-v3 wins.
Whisper’s best-documented failure also shows up in dictation. Its model card warns that it can output text that was never spoken and can repeat itself, and the Careless Whisper study linked hallucinations to long non-vocal pauses, which is exactly what happens when you stop to think mid-sentence. VAD filtering helps. Parakeet’s reported failures are different, mostly skipped sentences and wrong-language words, so neither model removes the need to read before you send.
IraVoice makes the same choice: NVIDIA Parakeet is the default, with Parakeet Unified EN (about 700 MB) downloaded on first run, and Whisper Small, Medium, Turbo and Large are heavier optional models you can switch to under Models when you need a language Parakeet does not cover. Hold Right Command on a Mac or Right Ctrl on Windows, speak, and release; the text lands at the cursor, and audio and text stay on the computer. IraVoice Local costs $79 once for one computer, $109 for two, or $149 for three after a 3-day free trial.