Dictāta Local: Private On-Device AI Dictation. Zero Cloud Dependencies.
Dictate into any application on Windows, Mac and Android without an internet connection. A streaming speech model runs on your own hardware, so your audio and dictated text stay on your device.
Time from stopping the recording to the final text, with streaming Nemotron 3.5. Measured on a Windows laptop CPU (Ryzen 7 5700U); similar in our tests on a MacBook with M1 (macOS 26) and a Google Pixel 9a.
Speech recognition and text cleanup run in local memory. While Local mode is on, the app blocks all connections except sign-in, billing, model downloads, the version check and feedback you send.
One streaming model on all three platforms, run through sherpa-onnx.
Your glossary terms are passed to the speech decoder as hotwords, without a language model.
Technical Architecture: How Dictāta Local Operates On-Device
Dictāta Local runs speech recognition and text cleanup on your own machine, in a separate background process, instead of streaming your audio to a remote server.
Local PCM Audio Capture
The microphone captures 16 kHz mono PCM audio in memory. On desktop, frames go to an isolated Electron utilityProcess through transferable buffers and are not sent over the network.
Streaming Neural Inference (sherpa-onnx)
Audio is fed in 100 ms chunks to NVIDIA Nemotron 3.5 ASR Streaming (0.6B, 560 ms, INT8), run by sherpa-onnx. The model recognizes speech while you talk, so the final text is ready about 120–200 ms after you stop.
Glossary Hints in the Decoder
Your Smart Glossary terms (up to 512) are passed to the decoder as hotwords with modified beam search, so names and jargon are favored during recognition. The usual text corrections run afterwards.
Deterministic Cleanup & Direct Insertion
Rule-based on-device cleanup removes fillers, puts paragraph breaks where you pause and applies voice commands and glossary corrections, then pastes the text at your cursor. Converting spoken numbers to digits is not supported yet.
Measured Performance
Numbers from our own measurements of the speech model that ships in Dictāta Local. We do not publish competitor figures we cannot verify.
| Metric | Dictāta Local | Basis |
|---|---|---|
| Stop-to-final text | ~120–200 ms | Windows 11, Ryzen 7 5700U, 4 CPU threads, recordings of 33–59 s fed at real-time speed (122 and 126 ms). The model alone, without paste into the target app. |
| Speech-model memory | ~0.8 GB | Process working set on the same machine; it stays flat regardless of dictation length. |
| Real-time factor | ~0.27 | Built-in speed check after download on the same machine. Lower is faster; below 1 keeps up with live speech. |
| Speech model | Nemotron 3.5 Streaming, INT8 | NVIDIA Nemotron 3.5 ASR Streaming 0.6B (560 ms), ONNX export from sherpa-onnx; about 475 MB download. |
| Languages | 28 | Pick a language in Settings or use automatic detection. |
Dictāta Local vs Superwhisper and Rhino Voice
A factual overview limited to what each vendor publishes. Details may change; check the vendor sites (superwhisper.com, rhinovoice.app) for current information.
| Feature / Spec | Dictāta Local | Superwhisper | Rhino Voice |
|---|---|---|---|
| Platforms | Windows, macOS, Android | macOS, Windows, iOS, Android | macOS 14+ on Apple Silicon (M1 or newer) |
| Speech engines | NVIDIA Nemotron 3.5 Streaming via sherpa-onnx | Local Whisper models; Parakeet and Cohere Transcribe local models and cloud models on Pro | On-device Whisper and Parakeet models, plus on-device text cleanup |
| Works without internet | Yes, after a one-time model download | Yes with local models; cloud models need a connection | Yes, after the models are downloaded |
System Requirements
- OS: Windows 10 / 11, 64-bit (x64 or ARM)
- CPU: Modern 64-bit processor; 4 or more threads recommended. Measurements above come from a Ryzen 7 5700U.
- RAM: 6 GB minimum, 8 GB recommended (the speech model uses about 0.8 GB while active)
- Disk: About 1.2 GB free for the one-time download and unpacking (the speech model is about 475 MB). Optional text models add 1.1–3.3 GB each.
- GPU: None required. Runs on the CPU.
- OS: macOS
- CPU: 64-bit processor, Apple Silicon or Intel
- Memory: 6 GB minimum, 8 GB recommended
- Disk: About 1.2 GB free for the one-time download and unpacking. Optional text models add 1.1–3.3 GB each.
- OS: Android phone or tablet
- RAM: Tested on a Google Pixel 9a; minimum requirements are being finalized during the preview
- Disk: One-time download of the speech model (about 475 MB)
Data Sovereignty: On-Device vs European Cloud
Some European dictation services advertise servers in Europe (for example Sprecho or Fast Dictate). EU hosting is a data-residency choice, but your audio is still sent to a server to be transcribed. In Dictāta Local the speech recognition runs on your computer.
When you select Dictāta Local, audio and dictated text are not sent anywhere, and usage analytics are blocked. The app denies every connection except sign-in and account details, billing pages you open, model downloads you start, the version check and feedback you send on purpose.
After the one-time model download, dictation works offline, for example on a plane or train. Signing in and checking your subscription still need a connection.
Audio and dictated text are not handed to any speech provider, so there is no external speech processor for dictation. Whether this meets your legal or professional obligations is for you or your compliance team to decide.
Frequently Asked Questions About Local Dictation
What is Dictāta Local?+
Dictāta Local is a mode of Dictāta that turns speech into text on your own computer or phone. It runs NVIDIA Nemotron 3.5 Streaming through sherpa-onnx, works without internet after a one-time model download, and does not send your audio or dictated text anywhere.
How fast is it?+
In our tests the final text is ready about 120–200 ms after you stop recording. We measured 122 ms and 126 ms on a Windows laptop (Ryzen 7 5700U) and saw similar speed on a Mac with an M1 chip (macOS 26) and a Google Pixel 9a. Older or heavily loaded computers will be slower. After the model download, Dictāta checks your machine with a short benchmark.
How does Dictāta Local differ from Superwhisper?+
Superwhisper offers local Whisper models, local Parakeet and Cohere Transcribe models on its Pro plan, and cloud models. Dictāta Local uses one streaming model, Nemotron 3.5, which recognizes speech while you talk. Check superwhisper.com for its current features and pricing.
How does Dictāta Local compare to Rhino Voice?+
Rhino Voice is a macOS app for Apple Silicon Macs (M1 or newer) on macOS 14 or later, with on-device Whisper and Parakeet models. Dictāta Local is available on Windows, macOS and Android.
Does Dictāta Local require a dedicated GPU?+
No. The speech model is INT8-quantized and runs on the CPU, using about 0.8 GB of memory while active.
Does it convert spoken numbers to digits?+
Not yet. Nemotron writes numbers as words, and we are working on converting them to digits.
Which features need Dictāta Cloud?+
Ask Dictāta chat, translation while dictating and Memory use Dictāta Cloud. Dictation, glossary, voice commands, history and, with an optional local text model, prompts, Casual / Formal styles and text actions work in Local.
Can I use Dictāta Local for confidential work?+
Audio and dictated text stay on your device while Local is on, and the app blocks other connections apart from sign-in, billing, model downloads and the version check. Whether that satisfies your legal or professional obligations is for you or your compliance team to judge.
Start Private Voice Typing on Your Device
Dictate into your applications with a speech model that runs on your own hardware.
Dictāta Local is in preview. It is included with the Lifetime license, any paid Dictāta Cloud plan and the free trial.