How Untapped turns speech into text
Untapped is a voice keyboard for Android by Sonicon Software. In any app where you can type, you switch to it, tap the mic and speak, and your words are typed in. A speech model you download once turns your voice into text on the phone, so your audio is never uploaded. It has no account, no ads and no tracking.
Here is what happens between tapping the mic and seeing your words.
1. Recording
Untapped records from the microphone only after you tap the mic. The sound goes into the phone's memory as 16 kHz audio, is fed to the speech engine on the phone, and is discarded when the dictation ends. It is never written to a file and never sent over the network.
2. Speech to text
A speech model you downloaded turns the audio into text on the phone, using the open-source sherpa-onnx engine. Streaming Zipformer models show words as you speak; Parakeet models transcribe after you stop and add capitals and punctuation. The model stays loaded for five minutes after you last use it, so the next dictation starts without a load wait.
On a Pixel 9 Pro XL, the multilingual model took about 6 seconds to load and finished a 30-second dictation 2.8 seconds after it ended.
3. Knowing when you stopped
A small voice-activity model, Silero VAD, which ships inside the app, listens for the end of your speech. After a pause of 0.8 seconds by default, a line under the text drains and the dictation finishes. You can set the pause from 0.3 to 2 seconds, turn auto-stop off, or tap stop yourself.
One dictation ends after 60 seconds at most, and the text so far is typed in.
4. Typing it in
When the dictation finishes, Untapped types the text into the field through Android's standard keyboard connection, adding a space if auto-space is on. An Undo chip appears for a moment so you can take it back. Like every keyboard, it reads the text in the field, here only to place spaces, undo and delete; it keeps none of it.
5. History
Finished dictations are saved in a private list on your phone so you can copy them again. Delete one, delete all, or untick Save dictations and nothing is written. Dictations in password fields and in fields that ask keyboards not to learn are never saved. Each entry holds the text, the time, how long you spoke and which model was used.
6. The models
| Model | In the app | Languages | Download | Live text | Licence |
|---|---|---|---|---|---|
| Parakeet TDT 0.6B v3 (int8) | Recommended, 25 languages | 25 European | about 639 MB | No, types after you stop | CC-BY-4.0 |
| Parakeet TDT 0.6B v2 (int8) | Text after you pause | English | about 630 MB | No, types after you stop | CC-BY-4.0 |
| Streaming Zipformer EN 20M (fp32) | Live text as you speak | English | about 87 MB | Yes | Apache-2.0 |
| Streaming Zipformer EN 20M (int8) | Smallest, live text as you speak | English | about 41 MB | Yes | Apache-2.0 |
The model list can be updated from the internet, but only models you haven't downloaded and not the recommended one can change. New models can also be added to the list.
Inputs and limits
- Dictation needs one speech model on the phone: about 41 MB for the smallest English model, about 639 MB for the recommended multilingual one.
- The recommended model understands 25 European languages; the other three models are English only. Hindi, Chinese, Japanese, Arabic and other languages outside that list are not supported.
- The streaming English models write in capital letters without punctuation.
- The Parakeet models type the text after you stop; only the streaming models show words while you speak.
- In password fields the keyboard says “Dictation is off in password fields. Switch keyboards to type here.”
- In fields where an app asks keyboards not to learn, such as an incognito tab, dictation works but nothing is saved to history.
- Untapped has no letter keys. Keep your usual keyboard for typing; the switch key takes you back to it.
- A single dictation ends after 60 seconds and the text so far is typed in.
- After the model unloads, the next dictation waits while it loads: about 6 seconds for the multilingual model on a Pixel 9 Pro XL, which also used about 1.6 GB of memory while transcribing. The model stays loaded for five minutes after you last use it.
- Runs on 64-bit Android phones with Android 8.0 or later.
- Check the text before you send it; speech recognition makes mistakes, especially with names, noise and accents.
What the internet is used for
Untapped uses the internet for two things. It downloads the speech models you accept from Hugging Face, asking first on a metered connection, and when you open the Models screen it may refresh the model list from GitHub. These requests carry none of your audio, text or settings, and dictation needs no connection.
The full details are in the Untapped privacy policy.
Speak instead of typing.
