Making one that works

Show somebody a still frame of a modern synthetic presenter and they will usually accept it. Show them eight seconds of the same presenter talking and a good proportion will say something is off, without being able to say what. That gap between the still and the motion is the entire craft problem, and it is a timing problem.

The viewer's eye is not running a resolution check. It is running a small set of expectation checks about how a speaking human behaves in time: does the mouth arrive with the sound, does the delivery breathe where a thought ends, does this body do anything at all while it talks, do these eyes ever close. Fail any of those and the verdict lands before the viewer has a word for it.

Which is why the honest advice on budget is unpopular. Higher resolution is the easiest thing to buy and the least of these four things.

It is not a click-path tutorial for any one product. Those go stale in months, the vendors publish their own, and two of the three products in this course were renamed in the last year.

It is also not a promise that craft is enough. This is the session where you get good at making one. Session 4 is where you find out what the evidence says about whether that helps you, and the answer is more uncomfortable than "yes".

Roughly 80 percent of what the vendor documentation teaches about capture and production is covered here. Certification and product training stay with the vendors, and Synthesia Academy is the only one of the three with a real certificate behind it.

Worth 18 points, the largest single contribution, and the one people most often skip because it looks like a script problem rather than a production problem.

What the viewer's eye is checking. Human speech is not evenly spaced. A comma is a small held breath, a full stop is a bigger one, and both of them are where the listener is given a moment to catch up. When those pauses are absent the delivery runs flat, and flat delivery is the fastest route to the uncanny feeling there is. The viewer reads it as a person who is not thinking about what they are saying, which is precisely what is happening.

What you actually do about it. Punctuate the script for delivery rather than for grammar. Put the comma where you would take the breath. Every one of the three vendors honours punctuation in its timing, so the script is your prosody control surface whether or not you think of it that way.

Worth 16 points. This is the one people think of as "lip sync", and it is the one that degrades most invisibly.

What the viewer's eye is checking. Whether this mouth belongs to these words. Not frame by frame, which nobody can do, but as an accumulating sense of drift. A generic mouth loop looks fine for a second and a half and then it is a syllable behind, and the viewer's verdict arrives around the point where a hard consonant lands on an open mouth shape.

What you actually do about it. Use the product that derives its mouth shapes from your specific audio rather than from a canned animation, and then respect the framing constraint the vendor publishes.

Read that as a shot-list rule. If your storyboard has the presenter in a wide corporate lobby shot, you have chosen the framing that most exposes the weakest part of the system.

Worth 14 and 12 points, and both are cheap. Together they are 26 points, more than lip sync, and neither costs a cent in resolution.

Blink, worth 14. The eye is checking that this face is alive at all. Blinking is involuntary, unconscious and constant, and its absence is one of the few things a viewer notices in under two seconds without knowing they have noticed it. A face that never blinks reads as wrong before anyone can articulate why. It does not need to be biologically accurate. It needs to happen.

Head motion, worth 12. The eye is checking for a stillness that no living body actually holds. Even a person concentrating on staying still micro-corrects constantly. Perfect stillness is the single strongest tell in the set, and small motion buys more believability than a resolution upgrade does.

All four levers work inside a system that is estimating rather than observing. Synthesia says this plainly about avatars derived from a photograph:

That is not a bug report, it is a description of the product class. A photo avatar has no performance behind it, so its motion is generic by construction. A video avatar has your actual performance behind it and behaves accordingly, which is the real reason the footage requirements in Part 3 exist.

Two more constraints worth knowing before you build on something. The avatar generation you pick has a shelf life: Synthesia's Legacy avatars "remain available... but are being phased out and are not recommended for new content". And the likeness can walk out with the person. On Synthesia, if the individual creates an account with the email address the avatar request went to, "the Personal Avatar will be transferred from your account to theirs".

A viseme is a mouth shape, the visual counterpart of a sound. There are far fewer of them than there are phonemes, because several sounds look identical on a face. The letters m, b and p all close the lips. That collapse is why lip reading is hard and why a mouth animation can be plausible without being right.

A viseme track is the ordered list of those shapes with a duration attached to each one. That duration list is the prosody. It carries the emphasis, the breathing, the sense that a thought ended, all of it. Change nothing but the durations and you change how human the delivery reads.

In the playground below, the track under the face is computed from the exact characters you typed. Vowels hold longer than stops, because they do: an open O is held around 115 milliseconds and a hard K around 70. A comma inserts 180 milliseconds of rest, a full stop 340, a space 60. Turn the prosody lever off and every punctuation pause collapses to 50, which is the flat delivery you have heard in every bad corporate explainer video.

The old received wisdom was that a synthesised voice is a learning penalty. Mayer's voice principle, human rather than machine narration, sits at a median effect of d = 0.78, which is substantial. But that finding is about the text-to-speech of its era.

Dincer 2022, with 51 participants, found human voice at d = 0.90 and modern text-to-speech at d = 0.75, both beating traditional synthesis, with no significant difference between human voice and modern machine voice at p = .200, and cognitive load statistically identical at p = .964. Craig and Schroeder found the same reversal.

The practical reading is narrow and useful. Modern synthesised speech is no longer the thing that gives you away. Timing is. If your presenter sounds wrong, look at the pauses before you look at the voice model.

Synthesia personal avatar. Between one and five minutes of footage, in a single continuous take, as .webm, .mp4 or .mov, under 2GB. The single-take requirement is the one that catches people out. You cannot stitch three good minutes together from six attempts, so plan the read as a performance rather than as a series of takes.

Language count, and why it is worth dating. Synthesia's documentation now says 31 languages. A figure of 33 circulates elsewhere. One of the two is stale, and rather than pick the more flattering one this course states the docs figure and the date it was read, which was 26 August 2026. Check it again before it goes in a proposal.

HeyGen Digital Twin. Minimum 1080p at 30 frames per second, with 4K at 60 preferred, in 16:9 or 9:16. This is the option that expects a real capture setup, and the resolution floor here is about giving the model something to learn a performance from rather than about output quality.

HeyGen Avatar V. A 15-second webcam recording, 175 or more languages, and no cap on output length. That is the shape of the current flagship: almost nothing at the door, and no limit on what comes out of it.

Avatar slot add-ons at HeyGen are $29 a month or $300 a year, each covering up to 500 looks for one individual.

There is a persistent story that D-ID was acquired and is winding down. It is backwards. D-ID is alive and was the acquirer, not the acquired. It bought the Berlin explainer-video company simpleshow, announced on 16 September 2025, and runs the combined company under the D-ID brand. Its own AI information page confirms that simpleshow is now part of D-ID.

Any vendor comparison that has quietly dropped D-ID is out of date, and any comparison that says D-ID was bought has the direction of the transaction wrong. This matters more than it looks: D-ID is the only one of the three where the watermark is never fully removable, even on Enterprise, which makes it a genuinely different option to have on a shortlist.

It also renamed its avatar tiers. "Express Avatar" is now V3 Instant, "Premium+" is now V3 Pro, and photo avatars are V2. The old names survive in D-ID's own help-centre URLs.

Craft does not move whether disclosure is required. That is decided by where you are publishing, what you are publishing about, and increasingly by statute. The UK advertising regulator's own position is that there is no blanket legal requirement to disclose AI in ads, and that the test is whether the audience would be misled without it. A better lip sync does not make an audience less misled. It makes them more misled, which is the wrong direction.

Craft does not move what happens when somebody finds out. An undisclosed synthetic presenter at 84 percent believability and one at 24 percent are in exactly the same position on the day someone works it out. The better one has simply been believed by more people for longer, which is more to unwind rather than less.

So take the 84. It is real, it is earned, and it is worth having. Then go into session 4 knowing that the axis you just mastered is the one nobody will hold you accountable for.