Guide · Offline speech to text
Offline speech to text that never uploads your voice
Offline speech to text runs the recognition model on your own computer. No audio is uploaded, no account is needed, and dictation keeps working on a plane, in a clinic, or in an air-gapped room. Here is how it works and how to set it up.
Updated 2026-11-11
one-time · GST included · FOUNDER BETA — FIRST 100 FREE
14-day money-back guarantee — if it doesn't hear you, email for a refund, no forms.
What offline speech to text actually means
Online dictation streams microphone audio to a server, transcribes it there, and sends text back. Offline speech to text does the same work locally: the microphone, the acoustic model, and the text output all stay on your machine.
Most offline apps still download a model once. That download is the only network request. After it finishes, the app can run with networking disabled and behave identically.
Why local recognition is worth it
Cloud dictation is convenient, but it makes a trade you may not want: every sentence goes to someone else's server. Local recognition removes that trade.
- Privacy — speech can contain patient notes, legal matters, source material, or credentials; local processing removes the upload.
- Latency — no round trip to a data centre, so the first words can land in under a second on a normal CPU.
- Reliability — dictation works on flights, in secure facilities, and during outages.
- Cost — no per-minute meter and no subscription; you pay once, or not at all during the beta.
How Saype does offline speech to text
Saype uses NVIDIA Nemotron 3.5 ASR Streaming 0.6B, a compact streaming model that runs on the CPU. It processes each new chunk of audio once and types interim results while you speak, then finalizes when you pause. That keeps latency low without a GPU.
A Whisper mode is included for tougher audio — large-v3-turbo accuracy and translation, still fully offline. Word boosting lets you add names, acronyms, and domain terms so they are recognized the first time.
The voice model is about 1 GB and downloads once. The app itself is a normal desktop installer for Linux, macOS, and Windows.
Offline vs cloud speech to text
Set up offline dictation in three steps
Saype is free for the first 100 founder-beta users and installs like a normal desktop app.
- Download the installer for your platform from the downloads page.
- Launch Saype and let it fetch the voice model once; after that it runs offline.
- Pick your hotkey, speak, and the transcript lands at the cursor in any app.
What accuracy should you expect?
On a clear microphone and ordinary speech, Saype's founder-measured first-time accuracy is 90%+; the rest is fixed live with Backspace revisions. Whisper mode targets 95%+ for difficult audio. Accuracy drops with heavy background noise, overlapping speakers, and unusual proper nouns — word boosting helps with the last one.
Treat any vendor's accuracy number as a starting point, not a guarantee. Test on your own microphone, your own vocabulary, and your own room.
Comparison at a glance
| Offline (Saype) | Cloud dictation | |
|---|---|---|
| Where audio goes | Stays on your machine | Uploaded to a vendor server |
| Internet needed | Only for the one-time model download | Every dictation session |
| Typical latency | Sub-second streaming on CPU | Network round trip plus processing |
| Account | None | Usually required |
| Pricing | One-time, no meter | Monthly subscription or per-minute |
| Works on a plane | Yes | No |
General comparison; exact behavior depends on the specific app and plan.
Frequently asked questions
Does offline speech to text need internet?
Only once, to download the speech model. After that Saype works with networking disabled.
Is offline speech to text accurate?
For clear speech, modern local models are close to cloud quality. Saype measures 90%+ first-time accuracy and offers a Whisper mode for harder audio.
Can I use offline dictation on a plane?
Yes. Once the model is downloaded, no network is required.
Does it add punctuation?
Yes. Saype produces punctuated, capitalized text and revises the tail with Backspace when the model updates a sentence.
Which model does Saype use?
NVIDIA Nemotron 3.5 ASR Streaming 0.6B by default, with an offline Whisper large-v3-turbo mode for difficult recordings.
How much disk space does it need?
About 1 GB for the default voice model, plus the app itself.