How does a voice activated teleprompter work, and why does it lose your place?

A voice activated teleprompter runs a speech recogniser on what you say, turns it into words, and matches those words against the script just around your current place; when it finds you, it scrolls that line to the reading position. It loses your place when the matching goes wrong: it jumps ahead when a common word or a repeated phrase matches somewhere further down, and it stalls when the recogniser hears something that is not in the script, or hears nothing at all.

Most complaints about voice scrolling are one of those two, and most are fixable, by the app or by you. Below is how, and how we built Orate’s voice activated teleprompter around them, limits included.

The three parts of voice scrolling

Every voice-following prompter, whatever it is called in the app, is three pieces in a row:

1. A recogniser. Speech recognition built into the phone, or a cloud service, listens to the microphone and produces words, revising its guess as it hears more.
2. A tracker. It takes each new word and asks where in the script you are. This is the part that decides whether the app is any good, and it is the part nobody advertises.
3. A scroller. It moves the text so the line you are saying sits where your eyes expect it, smoothly enough that you can keep reading.

Given the same stream of slightly wrong words from the recogniser, one tracker holds your place and another jumps a paragraph. That is where apps differ.

Why a voice teleprompter jumps ahead

The simplest tracker takes the word it just heard and looks for it in the script. That works for “Copenhagen” and fails for “the”. Short, common words appear on nearly every line, so a tracker that moves on one of them can be pulled to any line that happens to contain it.

Repeated phrases are the harder case. Scripts repeat themselves on purpose: a refrain, a product name, “in this video” at the start and the end. If the tracker searches the whole script, the second occurrence is as good a match as the first, and it can land on the wrong one. That is the jump people describe as the prompter “skipping a section”.

The third cause is the recogniser revising itself. It may write down one word, then change it a moment later once it has heard the rest of the sentence. A tracker that acts on every revision moves on guesses that were never final.

Why it stops scrolling

Stalling is the opposite failure: you are reading, and the text does not move. The usual causes:

• The recogniser wrote a different word. “Their” for “there”, a name spelled three ways, a number written as digits when the script spells it out. A tracker that only accepts exact matches sees no match and waits.
• Noise. A fan, traffic or a room with hard walls makes the recogniser less sure of every word, and less sure means more wrong words.
• Accent and language. Recognisers are better at some accents and languages than others, and some languages have no on-device recogniser on a given phone at all.
• You went off script. An aside, a correction, a cough. A tracker should stop here; the problem is when it cannot find you again afterwards.
• It cannot hear you. If something else has the microphone, the recogniser gets silence. During video recording this is the most common cause of all, and it has its own section below.

What a good tracker does differently

None of this needs a cleverer recogniser. It needs a tracker that is more careful about when to move:

• It matches a run of words, not one. Compare the last several words heard against the script, and a common word can no longer drag the text anywhere on its own. A long, distinctive word can be trusted sooner than a short one.
• It searches near your place first. The next few lines are far more likely than a line two pages away. Looking in a window around the cursor stops a repeated phrase further down from winning.
• It asks for more evidence the further it moves. One word to go to the next word; several in a row before it believes you jumped back or skipped ahead.
• It holds on off-script speech. When what you say matches nothing nearby, the text stays exactly where it is. Holding still is always better than guessing.
• It re-finds you on a restart. If you go back to the start of a sentence and say it again, that is the script, just behind the cursor, and it should follow you there.
• Commands cannot be mistaken for the script. If “go back” is a command, a script line that says “let’s go back to 2019” must still be read, not obeyed.

On-device or cloud recognition

Recognition runs either on the phone or on a server:

• Privacy. On device, nothing you say leaves the phone. With a cloud recogniser your script, read aloud, is sent somewhere else to be transcribed.
• Offline. On-device recognition works with no connection once the phone has its language model. A cloud recogniser needs a signal, which is a problem on location and in studios with poor reception.
• Delay. Every word sent to a server has to come back before the prompter can move. On device, there is no round trip.

The trade-off is language coverage: a phone can only recognise on device the languages it has a model for. That is why a voice teleprompter may follow you in one language and not another on the same phone. Test the language before the shoot, not during it.

The microphone while you record

One failure only shows up when you film. The encoder and the recogniser both need the microphone at the same moment, and two parts of an app that each ask for it do not always get to share it. When we tested in August 2026, PromptSmart’s VoiceTrack stopped following when we recorded video, and we wrote up the cause in why voice-tracked scrolling stops when you record video. That is a dated test of one app, not a claim about it today.

The fix is for the app to open the microphone once and give the same audio to both. That is how Orate works on iPhone: one capture session, with the same microphone buffers going to the recording and to the recogniser. We checked it on an iPhone 15 with a 78-second 4K take, and the audio in the file was intact.

Android is less tidy. The camera library Orate uses owns the microphone during a take, so the recogniser has to be a second listener, which relies on the concurrent audio capture Android 10 and later allow. It is built to do that, and we have not yet verified it on a real Android phone. What we did build is the fallback: if the phone silences the recogniser during a take, the recorder notices that the level meter shows speech while nothing has been recognised for eight seconds, and carries on at a fixed speed from where you are rather than stopping. The detail is in voice teleprompter and Android recording.

A checklist for reliable voice scrolling

Whatever app you use, these make the biggest difference, in rough order:

1. Test your language first. Read a paragraph with the camera off. If it does not follow you there, it will not follow you on the take.
2. A quiet room. Turn off fans and air conditioning for the take.
3. Keep the microphone close. A phone at arm’s length hears the room as much as you; a clip-on microphone gives the recogniser cleaner words.
4. Say the script’s words. A tracker can only follow words that are on the page. If you keep changing a line, change the script.
5. Write numbers and names the way you say them. It removes the most common mismatch between what the recogniser writes and what the script says.
6. Short lines and paragraphs. They give the tracker distinctive runs of words, and they give you somewhere to recover.
7. Record a short test take and play it back. Check the text followed you and that the sound is in the file.

If voice scrolling is not available for your language or phone, a calibrated fixed speed is the dependable alternative; teleprompter scroll speed shows how to find yours in twenty seconds.

How Orate’s Follow my voice handles it

In Orate, voice scrolling is called Follow my voice, an option in the recorder’s Speed sheet. Here is how it answers each problem above:

• Matching. Each word it hears is aligned with the five heard before it, against the script near your place: twelve words back to thirty ahead. Moving on to the next word needs one match if the word has four or more letters; a short skip forward needs two words in a row; a restart or a longer skip needs three.
• Recogniser spellings. Short words must match exactly. Longer ones match if they are within about a quarter of their length of the script word, so a near miss still counts.
• Off script. The text stops where it is. It never drifts ahead and never guesses, and silence holds it too. After three words that are not the script it is holding; after eight it is searching for where you came back in.
• Coming back. It picks up from the words you are saying, including when you restart a sentence or paragraph or skip a few lines ahead.
• Commands. Seven spoken commands, such as “go back”, “next paragraph”, “pause” and “start over”, in English, Spanish, Portuguese, French, German and Italian. A command is never confused with script text that contains the same words, and “stop” is deliberately not one. The full list is in teleprompter voice commands.
• On device only. It uses Apple’s and Google’s on-device recognisers. Nothing you say is sent anywhere, and it works offline once the phone has downloaded the language.

And the limits, plainly. Follow my voice is part of Orate Pro and is included in the free trial; the prompter and recording are free. It needs iOS 18 or later, or Android 12 or later, because Android 11 and older have no on-device recogniser. In a language the phone cannot recognise on device, the option is turned off with the reason shown, and the script scrolls at a fixed speed as always. Commands work only in the six languages above; in other languages it follows your reading but does not take commands. And commands are heard by the same microphone as the take, so they are in your recording, and you trim them out afterwards.

For the rest of the setup, from placement to text size, the pillar guide is how to use a teleprompter on your phone.

Questions people ask

How does a voice activated teleprompter work?

A speech recogniser turns what you say into words, a tracker matches those words against the script near your current place, and the prompter scrolls the matched line to the reading position. Faster or slower speech moves the text faster or slower.

Why does my voice teleprompter jump ahead?

Usually because it moved on a single common word, or found a repeated phrase further down the script. A better tracker matches a run of several words and searches near your current place before anywhere else.

Why does my voice teleprompter stop scrolling?

The recogniser wrote a word that does not match the script, noise or distance made it unsure, your language is not supported on the phone, or it cannot hear you because something else has the microphone, which is common while recording video.

Does voice scrolling work while recording video?

Only if the app gives the same microphone audio to the recording and the recogniser. Orate does this on iPhone. On Android it is built to and not yet verified on a real phone; if the phone will not share the microphone, it carries on at a fixed speed.

Is on-device speech recognition better for a teleprompter?

For most people, yes. Nothing you say leaves the phone, it works offline once the language is downloaded, and there is no round trip to a server. The limit is that it only covers languages the phone has a model for.

Is Orate’s Follow my voice free?

No. Follow my voice is part of Orate Pro and is included in the free trial. The teleprompter and recording are free, with a fixed scroll speed you set.