Make a Software Training Video With an AI Voiceover

What a Training Video Is For

A training video has one job. A new person watches it and then does the task on their own. Everything in this guide serves that job. Real screens, so what they see matches what they will click. A calm voice, so they can follow while they look. Captions, so they can watch with the sound off at the front desk. No music, because music fights the voice and makes steps harder to hear.

An AI voice makes this cheap enough to redo when the software changes. That matters more than the voice itself. A video you can rebuild in an hour stays true. A video that took a week to record goes stale and nobody fixes it.

Capture Real Screens With Playwright

Do not film a mockup. Do not film your own screen by hand, either. Hand recordings wobble, catch pop ups, and cannot be redone the same way twice.

Use Playwright to drive the real app and record it. Playwright is a free tool that opens a browser, clicks where you tell it, and saves a video of the tab. Because the steps are code, you can run them again after every update and get the same shots.

Set these up before you record:

• A fixed window size, like 1920 by 1080, so every clip lines up.
• A demo account with made up customers, jobs and numbers. Never film real customer data.
• Slow, human pacing. Add a short wait after each click so the viewer can see what changed.
• One clip per step. It is easier to match a clip to a line of script than to cut a long one.

Walk each flow once by hand first. Write down every click. That list becomes both your Playwright steps and the spine of your script.

Turn the Click List Into a Script

Each click on your list becomes one short paragraph of script. Start with why the step matters, then say what to click, then say what the viewer should now see. That last part is the one people leave out, and it is the one that tells a new person they did it right.

Here is what one step looks like when it is done.

When a customer calls to book, start a new job. Click Add Job at the top of the board. A blank job card opens on the right. Type the customer's name in the Customer name box. Then click Save. The card moves to the Booked column, and the customer gets a text with the time.

Notice what it does not do. It does not explain every field. It does not say how the feature was built. It does not use words like simply or just, which make a stuck viewer feel slow. It names the button, says what happens, and moves on.

Keep each video to one task. A video called How to Book a Job is one someone will find and finish. A video called Everything About the Board is one they will skip.

Write the Script for the Ear

A script that reads well on paper can sound strange out loud. Four rules fix most of it.

• Spell numbers as words. Write two hundred and twenty five, not 225. An AI voice may read digits one at a time, or read a year as a price. Words leave nothing to guess.
• Never write a statement shaped like a question. A line like so you click save, right, makes the voice rise at the end. It sounds unsure. Say you click save.
• Keep sentences short. One step per sentence. Say what to click and what happens next.
• Name what is on screen. Use the exact button label. If the button says Add Job, the voice says Add Job.

Read the script out loud once yourself. Any spot where you stumble is a spot the voice will stumble too.

Record One Continuous Take

This is the rule people break most. It is tempting to send each line to ElevenLabs on its own and join the clips. Do not. Each request starts fresh. The pitch resets, the pace shifts, and the breath between lines is wrong. Joined together, it sounds like a phone menu.

Send the whole script as one request. The voice then carries its tone and pace from line to line like a person would. Use a blank line between steps so the voice pauses there.

Use these settings. They give a calm, even read that suits teaching.

• Stability 0.75. Lower values wander in pitch. Higher values go flat.
• Style 0. Style adds drama, and training needs none.
• Speed 1.0. Slower sounds sleepy and faster loses new staff.

Pick one stock voice and keep it for every video in a series. A new voice in video three makes the set feel like it came from different places.

If the script is long, still send it as one take if your plan allows it. If you must split, split at a section break, never inside a step, and keep the same settings for every part.

Check Every Take With Whisper

AI voices misread things. They skip a word, say a number wrong, or say a brand name in a way no one would. You will not catch every slip by ear, because you already know what it should say.

So let a machine listen. Whisper is a free speech to text model. Run the take through Whisper and compare its transcript to your script, word by word. Every place they differ is a place to listen closely. Most are fine. Some are a misread you would have shipped.

When a take has a misread, fix the script, not the audio. Spell the word the way it sounds, or reword the line. Then record the whole take again. Patching one word into a take brings back the stitched sound you worked to avoid.

Add Captions and Put It Together

Whisper gives you timing for each phrase, so the same run that checks the take also gives you captions. Save them as an SRT file. Use your script's spelling in the captions, not Whisper's, so names and labels match the screen.

Then line up the clips with the voice. Each step's clip starts when the voice starts that step. If a clip is short, hold its last frame. If it is long, trim the waits. Burn in the captions or ship the SRT beside the video, based on where it will play.

Do not add music. Not soft music, not intro music. It makes a quiet voice harder to hear and adds nothing a learner needs.

What the Finishing Pieces Contain

You can make a solid training video from the steps above. The finishing pieces give you the code: a Playwright capture script with pacing built in, the ElevenLabs request with these exact settings, and a Whisper check that prints every word that differs and writes the caption file. They also cover the misreads that bite most often, and the checklist we run before a video goes to staff. Next, see the whole pipeline in our training film build.