Lesson 5 of 8 · 8 min read

Voice, avatar or screen recording

Compare a human voice, an AI voice and an avatar platform on sound, consent and cost, with prices read on the vendors’ pages on October 15, 2026.

Three routes to a finished video

Lesson 1 decided whether your video needs a presenter. This lesson turns that decision into tools and a budget. Most workplace training videos take one of three routes. Each route has its own cost, its own consent work and its own cost of updates.

The first route is a screen recording narrated by a person. You record the screen, then record your own voice or a colleague’s. It costs the least and needs no AI. Its weak point is retakes, because every script change means another recording session.

The second route is a screen recording with an AI voice. You record the screen, and a text-to-speech tool reads the locked script from Lesson 3. ElevenLabs sells AI voices on their own. Descript, a video editor that works from the transcript, includes AI voices too. A script change means editing a line of text and generating the audio again.

The third route is an avatar platform. Synthesia, HeyGen and Colossyan generate a presenter who reads your script on screen. They suit content that changes often or needs many languages. They also bring the consent and disclosure work that Lesson 6 covers.

Judge the voice first

Audio quality matters more than whether the voice is human or AI. TechSmith, which sells the Camtasia screen recorder, tested narration with 768 full-time workers in four countries. Quiz scores rose as voice quality rose. Many viewers could not tell a high-quality AI voice from a human one, and robotic audio hurt ratings.

Read that study with its caveats. TechSmith studied its own product category and published the results on its own blog. Its avatar study gives subgroup counts of 36 and 32 viewers, with no overall sample size.

Older advice says to avoid synthetic voices. It rests on Richard Mayer’s voice principle, which drew on three experiments from 2003 and 2005. The synthetic voices in those experiments are two decades old. In 2017, Scotty Craig and Noah Schroeder compared a classic voice engine, a modern one and a human voice.

The modern engine produced more learning on transfer tests than the classic one. Learners rated it as high as the human voice for credibility. The authors said this calls the older findings into question.

So choose a voice by listening, not by its label. Before you pay, generate a 30-second sample that reads every term on the pronunciation list you started in Lesson 3. Product codes, supplier names and acronyms are where voices fail. User reviews of avatar tools name pronunciation as a common complaint.

What an avatar adds, and what it does not

Lesson 1 covered the avatar studies, and they bear on this choice too. In Torbjørn Netland’s study of 447 online participants, exam results showed no significant difference between human-made and AI-made videos. Learners enjoyed the human-made videos more. A study of 108 undergraduates found no performance difference and lower engagement with the AI instructor.

Both used short videos and tests soon after viewing. Neither measured what people did at work. So an avatar is acceptable for basic knowledge transfer. Neither study shows that an avatar improves learning or engagement.

If you use an avatar, keep it small. In TechSmith’s avatar study, viewers rated a picture-in-picture avatar higher than a full-screen one. Full screen made robotic traits easier to spot. In a procedure, the screen carries the lesson, so the avatar should not cover it.

What the tools cost on October 15, 2026

Emeri read each vendor’s own pricing page on October 15, 2026. Plans change often, so open the live page before you buy. Colossyan repackaged its plans in July 2026, and third-party price articles often lag behind the vendors’ pages.

Synthesia has a free Basic plan, Starter at $29 a month or $18 a month billed yearly, and Pro at $89 a month or $64 billed yearly. Enterprise is priced on request. SCORM export and 1-click translation into 140+ languages appear only on the Enterprise card.

SCORM is a package format that a learning management system (LMS) can import and track. HeyGen has a free plan, Creator at $29 a month, Pro at $49, and Business at $149 plus $20 per seat. SCORM export and LMS integrations start at Business. Proofreading a translated script starts at Pro.

Colossyan’s Starter plan is free, with no SCORM export. Professional costs $59 a month billed yearly, plus $30 a month for each extra member, and includes 5 SCORM exports a month. Enterprise is priced on request, with unlimited exports.

ElevenLabs lists Starter at $6 a month, Creator at $22 and Pro at $99. Starter includes about 30 minutes of speech, a commercial license and instant voice cloning. On October 15, Starter showed a $1 first month until October 18, and Creator an $11 first month. Cloning a real person’s voice needs the consent work in Lesson 6.

Descript lists Hobbyist at $16 and $24 per person a month, Creator at $24 and $35, and Business at $50 and $65. The page does not say which figure is the annual rate. AI voices, Studio Sound and filler-word removal start at Hobbyist. Translation proofreading starts at Business.

SCORM export is the dividing line for a small budget. It needs Synthesia Enterprise, HeyGen Business or Colossyan Professional. Without it, export an MP4 and place it on a course page in your LMS, or wrap it in an authoring tool. Lesson 7 covers that step.

Generated footage, and tools that have gone

Generated video models, such as Google’s Veo 3.1 and the models Runway sells, make short clips from a text prompt. Lesson 4 described what learners reported from such clips: forklifts that merged into each other, and garbled text on documents. Use them, if at all, for a short scene-setting shot with no text. Never use them to show a procedure, a screen or a safety step.

Two names in older guides are out of date. OpenAI shut down Sora in 2026. The app closed on April 26 and the API on September 24. Wix acquired Hour One in May 2025. Treat any tool list from before 2026 as a list to check, not a buying guide.

Worked example · Larkfield Supply

Riley prices two 4-minute segments

Lesson 1 chose a narrated screen recording with an AI voice and no avatar. Riley still prices all three routes, because Riley’s manager asked about an avatar of Morgan. The video is two segments of about 4 minutes each, for 25 staff. Larkfield’s LMS accepts an MP4 on a course page.

  1. Route one, Riley’s own voice, costs nothing extra. Riley records a test paragraph at a desk and hears the warehouse radio behind it. One sentence takes three attempts. Every later script change would mean another session.
  2. Route two needs a voice and an editor. ElevenLabs Starter, at $6, covers about 30 minutes of speech, which leaves room for retakes on 8 minutes of script. Descript Hobbyist adds editing and its own AI voices. Riley budgets $24 for one month, because the page does not label the monthly price.
  3. Riley generates the pronunciation list from Lesson 3 in both voices. One voice reads “LKS” as a word. The other puts the stress in “Aubriot” on the wrong syllable. Riley adds the respellings from the list, “L-K-S” and “oh-bree-OH”, and generates both samples again.That answers the question you guessed at: Guess: where do AI voices most often go wrong in a training script?
    What would you do next?Both voices now pass the list. How should Riley choose between them?Think it through, then show the answer

    By listening to a full minute of the script in each voice, with two colleagues who will take the training. The cost difference is $6 for the month. Voice quality matters more to learners than that difference.

  4. Route three would need an avatar platform. Synthesia Starter, at $29 a month, includes up to about 12 minutes of video, enough for 8 minutes. It has no SCORM export. The cheapest SCORM route is Colossyan Professional at $59 a month billed yearly, a commitment of $708.
  5. Riley adds the costs that no pricing page shows. Morgan would spend time recording the vendor’s consent video and reviewing a release. Riley would also generate a new avatar video every time the returns screen changes.
    What would you do next?Does Riley need SCORM for this video?Think it through, then show the answer

    No. The LMS plays an MP4 on a course page, and Lesson 8 checks learning with scenario questions, not with a completion flag. SCORM would add cost without adding evidence of learning.

Riley chooses route two, at about $30 for the first month, and writes the tools, the price and the date read into the video plan. The avatar would have cost more in money and in consent work. The studies give no reason to expect better learning from it.

Starter prompts

  • Compare production routes

    I am planning a training video. Here is my plan: [paste the objective, the script length in minutes, the number of learners, how often the process changes, the languages needed, and whether the video has a presenter]. Here are the prices and plan limits I read today on each vendor's own pricing page, with the date: [paste them]. Compare three routes: screen recording with a human voice, screen recording with an AI voice, and an avatar platform. For each route, give the cost for one month, the work each script change needs, whether I can export SCORM, and the consent work involved. Use only the prices I gave you. If a price or limit is missing, write "check the live page". Do not recommend a vendor I did not list.

  • Build a pronunciation test script

    Here is my pronunciation list: [paste each term and how it should sound]. Write a 30-second test script for a text-to-speech voice. Use every term at least twice: once at the start of a sentence and once in the middle. For each term, suggest a respelling that a voice tool is less likely to misread, and mark each suggestion as a guess for me to test by ear. Do not add terms I did not give you.

When not to use AI

  • Generated footage for any shot that shows a procedure, a screen, a document or a safety step. Text and physical detail come out wrong.
  • An avatar or a cloned voice of a colleague before the consent work in Lesson 6 is done.
  • Choosing a voice from a vendor demo. Test it on your own pronunciation list and script.
Recall from memoryHow do you choose an AI voice before you pay for it?Answer it without looking back, then show the answer

By listening, not by its label. Generate a 30-second sample that reads every term on your pronunciation list, because product codes, supplier names and acronyms are where voices fail.

Exercise · your project

Choose a production route for your video and price it on today’s pages.

  1. Write down your script length in minutes, your number of learners, and how often the process changes.
  2. Open the live pricing page for each tool you are considering, and note each price and plan limit with today’s date.
  3. Generate a 30-second sample of your pronunciation list in each voice you are considering, and listen to each one.
  4. Decide whether you need SCORM export, or whether an MP4 in your LMS will do.
  5. Add the route, the tools, the monthly cost and the date you read the prices to your video plan.

Sources