Skip to content

6 min read Web Dev Career

I Can't Memorize Lines, So I Built a Teleprompter That Listens

Blog FollowSpot is a voice-following browser teleprompter that runs Whisper locally. Why I built it, plus benchmarks of five Whisper models on an M5 Max.

TL;DR I can't memorize lines, so I built FollowSpot, a browser teleprompter that follows my voice using a Whisper model running locally through whisper.cpp. It lets me pick the model and set up the screen exactly how I want, including a narrow column centered over the camera lens. I benchmarked five Whisper models on my M5 Max MacBook Pro with my own voice: every one kept up, tiny.en mangled tech jargon, base.en missed one more term than the bigger models, and small.en and up got 14 of 16 right.

On this page
  1. Elgato already follows your voice. I wanted to pick the model
  2. Put the words over the lens, or everyone can tell you’re reading
  3. How it keeps up when you go off script
  4. I benchmarked five Whisper models. On an M5 Max, all of them keep up
  5. What’s still rough
  6. Try it

I have the energy of a theater kid and I can’t memorize lines. I’m pretty sure something is wrong with my brain. (If anyone knows if this is a real medical disorder, please let me know). I can write a paragraph, read it twice, and lose it by the time the camera is rolling. It’s the reason I never had the guts to do theater.

Public speaking worked out anyway. In DevRelApr 21, 2026Running DevRel in 2027 Means Owning Three Discovery Surfaces I mostly talk about things I built, so there’s nothing to memorize. I just explain the cool thing I’m working on.

Lately I’ve been making more videos, though, and some of them need tight monologues where every sentence has a job. Winging it doesn’t work there. So I built FollowSpot, a browser teleprompter that listens while I read and follows along. The speech recognition is a whisper.cpp model running on my own laptop, and my audio never leaves it.

Elgato already follows your voice. I wanted to pick the model

Voice-following teleprompters exist. Elgato’s Camera Hub has one called Voice Sync, and it runs on-device, which I appreciate. You just can’t choose what’s listening. Elgato doesn’t say which model Voice Sync uses. One developer who dug through the Mac app reports that it bundles whisper.cpp with base.en, one of the smallest Whisper models.

Pawn Stars Best I Can Do meme: Rick Harrison hears 'I'd like to pick my own Whisper model' and answers 'Best I can do is base.en'.

Browser teleprompters are worse on the privacy front. Many of the ones that follow your voice use Chrome’s Web Speech API, which sends your audio to Google unless the site has opted into Chrome’s newer on-device mode.

They Don't Know meme: a guy in a party hat stands alone in the corner while everyone else dances, thinking 'They don't know their browser teleprompter is sending their audio to Google'.

Meanwhile my MacBook Pro has an M5 Max in it. It can run models far bigger than base.en in real time, and in my experience a bigger local model is more accurate than anything built into a prompter app. I also had a specific picture of how I wanted the screen to look, and nothing let me set it up exactly that way. So FollowSpot takes any whisper.cpp model you point it at, from tiny.en up to large-v3-turbo.

Day to day I run medium, and I didn’t even download it. Screen Studio ships a copy, and FollowSpot picks it up automatically.

Put the words over the lens, or everyone can tell you’re reading

Keep the text narrow and centered right over the camera lens. I wish someone had told me that years ago. With a wide column, your eyes sweep left to right on every line, and viewers notice that little side-to-side scan. StudioKitGuide’s eye-line guide gives the same advice. Here’s how the prompter is mounted in front of my cameraNov 9, 2026A tour of my recording setup in 2026, all on one standing desk that rolls.

This is what a default wide column looks like:

FollowSpot with a wide text column stretching most of the way across the screen, so each line of the bread recipe script runs about ten words

And this is how I run it, with the column at 30% of the screen, top and bottom margins, and the reading line nudged up to lens height:

FollowSpot with a narrow text column in the center of the screen, three or four words per line, so the eyes barely move while reading

Three or four words per line means my eyes barely move. All of it is a slider in the toolbar, saved in the browser. On a beam-splitter prompter like Elgato’s, I drag the window to the prompter’s display, go fullscreen, and hit M to mirror the text.

FollowSpot's toolbar at the bottom of the screen with Listen, text size, width, top margin, bottom margin, and reading line sliders, plus Mirror and Fullscreen buttons

How it keeps up when you go off script

While I’m talking, the page keeps sending the last few seconds of audio to whisper.cpp on 127.0.0.1 and fuzzy-matches what comes back against the script near my place. A flubbed word or an ad-lib doesn’t throw it off, and when I stop talking, it stops. If I blow a take, I click the word I want to restart from.

I record those takes with Screencast Director, a screen recorder I built that drafts the edit for meOct 12, 2026I hate editing video, so I built a screen recorder that edits for me. When FollowSpot is open, it saves the script with the recording and keeps only the last take of each line.

Scripts are plain Markdown, and stage directions in *[brackets]* show up dimmed so I don’t read them out loud by accident.

I benchmarked five Whisper models. On an M5 Max, all of them keep up

I wanted real numbers, so I read a paragraph from my homelab postMar 5, 2026My Homelab Two Years Later, Every Upgrade and What It Cost out loud: 86 seconds, full of words like Proxmox, ThinkServer, Frigate, and VLANs. I ad-libbed a few words along the way, which turned out to be a good test by accident. Then I ran that recording through every model, using FollowSpot’s own matching code and the same listen loop the browser runs.

ModelSizeRAMLatency p50 / p90Highlight lag p50Tech terms right (my voice)WER, synthetic voices (clean / noisy)
tiny.en75 MB273 MB28 / 39 ms217 ms9 of 1613.0% / 15.8%
base.en142 MB373 MB33 / 44 ms172 ms13 of 167.9% / 12.2%
small.en466 MB813 MB66 / 110 ms192 ms14 of 167.2% / 7.8%
medium1.5 GB2.0 GB147 / 272 ms431 ms14 of 167.8% / 9.4%
large-v3-turbo1.6 GB1.9 GB184 / 228 ms249 ms14 of 169.2% / 10.1%

Latency is one request to whisper.cpp. Highlight lag is how long after I finished saying a word it got marked as read. Every model followed 100% of the script and never lost my place.

Tech terms heard right, of 16 0 8 16 tiny base small medium turbo 9 13 14 14 14 Median latency, M5 Max 0 100 200 tiny base small medium turbo browser's 250 ms tick 28 ms 33 ms 66 ms 147 ms 184 ms
Tech terms heard right, of 16 0 8 16 tiny base small medium turbo 9 13 14 14 14 Median latency, M5 Max 0 100 200 tiny base small medium turbo browser's 250 ms tick 28 ms 33 ms 66 ms 147 ms 184 ms

Speed isn’t the constraint on this laptop. Even large-v3-turbo answers in about 180 ms, which fits inside the browser’s quarter-second tick. On an older Intel machine that would be a different story.

The smallest model falls apart on jargon. tiny.en heard “Xeon E3-1226” as “Zion E3, 1, 2, 2, 6” and “ThinkServer” as “things server.” base.en fixed the Xeon and still wrote “the thing server.” Everything from small.en up got ThinkServer right. And every single model wrote “R stack” for the arr stack, which, fair.

Mocking SpongeBob meme: the top reads 'Me: ThinkServer' and the bottom, over a mocking SpongeBob, reads 'tiny.en: tHiNgS sErVeR'.

Above small.en, I couldn’t measure a difference. On this script, small.en, medium, and large-v3-turbo all landed on 14 of 16 terms. To get word error rates I also ran four synthetic macOS voices, clean and with pink noise at 15 dB SNR, and those flattened out at the same point. The leftover errors were mostly the synthetic voices mispronouncing things (“Flex” for Plex) in ways no model could fix.

On an Apple Silicon Mac, run small.en or bigger, because the latency costs you nothing. On an older or CPU-only laptop, small.en is still the one to try first; add the -ac 512 flag from the README, which cut CPU latency about 4x in my testing. Skip tiny.en unless you’re debugging.

I stick with medium because it’s already on my machine. small.en would do the same job on this script for less than half the RAM.

What’s still rough

It’s Mac-first. It runs on Linux if you build whisper-server from source, and I haven’t tried it on Windows at all. The launcher pins Whisper to English right now. And every number above comes from one paragraph read once on one laptop, so run it on your own scripts before you trust mine.

Try it

FollowSpot is on GitHub at JoeKarlsson/followspot, MIT licensed:

brew install whisper-cpp
git clone https://github.com/JoeKarlsson/followspot.git && cd followspot
./followspot download small.en
./followspot download vad
./followspot your-script.md

Press Space, allow the microphone, and start reading.

Works on My Machine

Get the next post by email. My newsletter covers homelab builds, AI tools and open source, and it goes out when I have something to say.

It's free, and you can unsubscribe anytime. You'll get an email to confirm first.

If you liked this post, consider sharing it on Hacker News, Bluesky, LinkedIn, or X.

Joe Karlsson at his desk, chin on hand, with a microphone in front of him

Joe
Karlsson

I write about my homelab, AI tools, databases, and the weird stuff I build on the side.

More about me →

BlueskyLinkedInGitHub