How to type by voice in any Android app, with the speech recognised on the phone

Dictation in any app, from a floating bubble

The point of dictation is that it works where you already are, not in a box you copy out of. On Android, Deixis gives you a floating microphone bubble that dictates into any app's text field: a message, an email, a note, a search box. You stay in the app you were using, and your words land in the field you were typing in.

The words are recognised on the phone. The speech model runs on the device, so what you say is turned into text there rather than sent off to be transcribed. It is the same engine as the desktop builds: Deixis is one engine with four hosts — Windows, Mac, iPhone and Android — and recognition happens on the device on every one of them.

To get it, install Deixis from Google Play. The site's download link, opened on an Android phone, takes you there.

Meetings, recorded and transcribed on the phone

The Android app also records meetings, from the phone's own microphone, and transcribes them on the phone as they run, so the transcript is building while the meeting is still going rather than arriving afterwards. No bot joins anything: it is your phone, listening through its own microphone.

Notes can be written from the transcript. They are one of the optional cloud features: the transcription happens on the phone, and writing notes is the step that asks a language model, which is sent the transcript and never the audio. The meetings page describes meeting mode as it works on Windows, where it also records the other side of a call from the system audio when it can.

What carried over from the desktop

Alongside the bubble and meetings, these came across from the desktop app:

• Translate, one of the optional cloud features, for text you want in another language.
• The Library, where your history is kept.
• The dictionary, so you can teach Deixis the names and jargon you use.
• Your account, which is what the managed cloud features and the trial hang off. Deixis does not ask you to sign in to dictate.
• Both speech models: the standard one, and a larger, more accurate one that is slower to decode.

What stays on the phone, and what can leave

Speech recognition runs on the device, and your audio is never uploaded — not a dictation, not a meeting, under any setting or on any plan. The optional cloud features — cleanup, translation, meeting notes and a suggested meeting name — are the ones that send text to a model, and they never send audio.

When one of those does run, what it sends is text, never audio:

• Cleanup sends the transcript of what you just said and your vocabulary terms, with some context about the field it is going into, so the model can punctuate it for where it is going — and, only if you switch on Send surrounding text, the 400 characters before the cursor.
• Translate sends the text you asked it to translate, and your vocabulary terms.
• Meeting notes send the meeting's transcript.
• Meeting names, when you press Suggest one or switch on Better meeting names, send the opening of the transcript.

Cloud text sent through Deixis's own endpoint is routed only to model endpoints that keep no copy of it and do not use it for training. The full list of what leaves, row by row and with the exact conditions, is on what leaves your machine. The platforms page has the current state of each build.

Getting your own words right

Out of the box a speech model does not know your colleagues' names, your products or your field's jargon. The dictionary is the biggest accuracy lever there is, and it costs next to nothing in speed. It holds:

• Words to expect — names and terms the recogniser should listen for.
• Fixes — corrections applied after recognition, at word boundaries, for the words it keeps getting wrong. Write the correction once and it stays corrected.
• Filler words — extra words to strip, beyond the built-in "um" and "uh".

Words to expect bias the recognition itself, before any cleanup runs, which is why a misheard name is the dictionary's job rather than cleanup's: cleanup rewrites prose, not hearing. The dictionary page explains how it works, including the cap on how many terms reach the speech model and why fixes have no cap.

If you dictate long passages, the larger speech model is the next lever, at the cost of a slower decode. And no setting recovers a word the microphone never caught, so in a noisy room it is worth speaking closer to the phone.

What the Android app does not do

Worth knowing before you install rather than after.

• Meetings come from the phone's microphone. On Windows, Deixis can also record the other side of a call from the system audio; the Android app records what the phone's microphone hears.
• No screen grabs, screen recording or free mode. Those are desktop features, on Windows and the Mac.
• English speech models. Both models that ship are English ones.

And if you use an iPhone: the Deixis keyboard for the iPhone is in beta on TestFlight. It is not in the App Store.