We build AirTypes, a desktop dictation app with both sides of this story: Plain mode is free and runs Whisper offline on your CPU, and the optional Polished mode cleans up your words in the cloud, 120 minutes of speech for every dollar. Here is the same messy sentence through each:
"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"
The meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again.
Polished, which sends the recording to AirTypes, typed: "The meeting is moving to tomorrow at 10:00. The client is not available. Send him the invoice again." The offline version gets every word right; the cloud version also rewrites. That gap is the real trade-off in 2026, and the rest of this piece explains why.
What offline speech recognition is
Offline speech recognition converts spoken audio into text entirely on your own device. A neural network runs on your CPU (or GPU), takes your microphone audio, and outputs text, with no round trip to a server. Because nothing is uploaded, it works on a plane, in a secure facility, or on a flaky connection, and the audio stays on the machine.
This is the opposite of cloud STT (Google Speech-to-Text, AWS Transcribe, Otter, Wispr Flow and most "voice typing" services), where your audio is sent to a vendor's servers for processing.
How on-device ASR works
At a high level, every modern speech recognizer follows the same pipeline:
- Capture. The microphone records raw audio (typically 16 kHz mono for speech).
- Feature extraction. Audio is converted into a spectrogram: a numeric picture of frequencies over time.
- Acoustic and language modeling. A neural network maps the spectrogram to text. Whisper, for example, is a transformer trained on a large multilingual dataset that does this end to end, handling punctuation and casing too.
- Decoding. The model produces the most likely text sequence, optionally with timestamps.
The key point for "offline": all four steps run from a model file already on disk. Once downloaded, the model needs no network. AirTypes Plain downloads the Whisper model you pick once, then runs this pipeline locally every time you speak.
Offline vs cloud speech-to-text
| Dimension | Offline (on-device) | Cloud |
|---|---|---|
| Privacy | Audio stays on the device | Audio sent to the vendor |
| Works without internet | Yes, always | No |
| Latency | Local, depends on your hardware | Network round trip |
| Ongoing cost | Free after model download | Per minute or per month |
| Accuracy | High at the larger model sizes | High |
| Cleanup and rewriting | Light, rule-based | Can use a large language model |
| Languages | Depends on the model and app | Usually many |
| Hardware needed | Your CPU/GPU and RAM | None (vendor's servers) |
The historical knock on offline, worse accuracy, no longer holds at the larger model sizes. Two trade-offs remain. You spend local compute and RAM instead of money per minute. And heavy cleanup, turning rambling speech into tidy sentences, usually needs a large language model, which is too big for most laptops and so tends to live in the cloud.
That is why some tools now offer both: offline transcription by default, and a cloud step you choose when you want it. Superwhisper's Pro plan adds cloud models alongside local ones; AirTypes has Plain (offline) and Polished (cloud). What matters is that you know which one is running.
The offline ASR engine landscape
- OpenAI Whisper: the dominant open model. Strong accuracy, around 99 languages in the multilingual models, robust to noise and accents. The default choice in 2026.
- whisper.cpp: a fast C/C++ port of Whisper that runs efficiently on CPU and Apple Silicon. It is what makes Whisper practical in lightweight desktop apps.
- faster-whisper: a CTranslate2-based reimplementation that is faster and lighter on memory, popular for near-real-time use.
- Vosk: a compact, low-latency toolkit that runs on tiny devices (even Raspberry Pi and phones). A lower ceiling than Whisper, but great for embedded and streaming.
- Coqui STT / DeepSpeech: older open engines, now largely succeeded. Still seen in legacy projects.
- NVIDIA NeMo / Parakeet: high-accuracy models for people with NVIDIA GPUs and patience for heavier setup.
For a desktop dictation workflow, Whisper (via whisper.cpp or faster-whisper) is almost always the right engine. Vosk wins when you need very low latency or minimal hardware.
Accuracy vs model size
Whisper ships in sizes, and the choice is a direct speed-vs-accuracy dial:
- tiny / base (about 78 to 148 MB): fastest, lowest RAM, more errors on names and jargon. Good for quick notes on weak hardware.
- small / medium (about 488 MB to 1.5 GB): the sweet spot for many laptops: strong accuracy, reasonable speed.
- large-v3-turbo (about 1.6 GB): close to the top of the family at a much higher speed than full large-v3.
- large-v3 (about 3.1 GB): the largest and slowest member of the family.
Practical rule when you run Whisper yourself: pick the largest model your machine runs comfortably. AirTypes Plain keeps to the two smallest English models, tiny as Swift and base as Crisp, and Crisp is the tier we recommend. When you need more accuracy than Crisp gives, Polished transcribes with a large speech model on the server instead. The full breakdown is in our Whisper model sizes guide.
When offline is the right choice
- Privacy and compliance: legal, medical, finance and government work where audio should not go to a third party. See our enterprise offline voice recognition guide and is voice dictation private?
- No or poor connectivity: travel, fieldwork, air-gapped systems.
- Cost at scale: heavy daily dictation where per-minute cloud pricing adds up.
- Latency control: no network variance; speed depends only on your hardware.
Cloud makes sense in the other cases: when you want finished, grammatical text without editing, when you dictate in a language your offline model does not cover, or when your laptop is too slow for a good local model.
From an engine to a usable app
An engine is not a workflow. Whisper transcribes a file; it does not capture your microphone on a hotkey, clean up filler words and type text at your cursor in any application. That last mile is what a dictation app provides.
AirTypes wraps Whisper into exactly that on macOS and Linux (the Windows desktop app is in development). Hold Ctrl + Shift + Space, speak, let go, and the text is typed at your cursor. In Plain mode the whole pipeline runs on your machine, free, in English. Switch to Polished on the Home screen when you want cloud cleanup in your language: Polished sends the recording to AirTypes to transcribe and clean it up, and it is prepaid ($5 buys 600 minutes, credit never expires). See how Polished works. On Windows today you can Dictate in your browser (new in November 2026), but that is Polished only: the recording is sent to AirTypes, so it is not offline.
- Want to get set up? Start with the practical offline voice-to-text guide.
- Want to compare finished apps? See the best offline speech recognition apps comparison.
- On Linux? See the complete Linux voice dictation guide.
FAQ
What is offline speech recognition?
It converts speech to text entirely on your device, with no audio sent to a server. A neural model runs locally, so it works with no internet and keeps audio on your machine. It is also called on-device ASR or offline STT.
Is offline speech recognition accurate?
Yes. Modern open models like Whisper are accurate enough for daily dictation, and accuracy scales with model size: smaller models are faster but make more mistakes on names and jargon, larger models are more accurate but slower and need more RAM. What offline tools usually lack is heavy cleanup, which tends to need a large language model in the cloud.
What is the best offline speech recognition engine?
For most uses, Whisper (via whisper.cpp or faster-whisper) gives the best accuracy for the effort. Vosk is best for lightweight, low-latency or embedded scenarios. The right engine depends on your hardware, latency needs and languages.
Does Whisper work offline?
Yes. Once the model file is downloaded, Whisper runs fully offline. AirTypes Plain mode runs Whisper on your CPU this way, so transcription happens locally with no audio uploaded.
Get AirTypes, free
Plain runs Whisper offline on your computer, free with no time limit. Polished is optional when you want cleaner text in your language. The desktop app runs on macOS and Linux now; the Windows app is in development.
Download for macOS and Linux