Digital avatar for a brand: where implementation breaks down
Initial launches of a talking digital presenter are convincing in a five-second pilot clip and fall apart in a one-minute working piece.
There it becomes visible that articulation lags behind the voice, the lower jaw lives separately from the lips, and the gaze does not reach the camera — and the viewer reads the fake before they manage to understand what is wrong.
The cause is almost never the quality of generation as such. A talking head is assembled from two independent streams: speech is synthesized by phonemes, while the face is animated by sound volume, so sibilants and plosive consonants systematically fail.
On top of that, facial layers conflict — eyebrows, eyelids, and mouth corners are controlled by different rules and together produce a “plastic” face.
The second source of failures is the tone of delivery. The voice is reconstructed in separate sessions, each with its own emotional background: the first line sounds composed, the tenth neutral, and the integration with the same text comes out deliberately cheerful.
A unified image of the speaker falls apart — lighting, head position, tempo, and stress placement change.
Third — the production cycle. While there is no managed pipeline, any text edit means reassembly: reference recording, re-synthesis, redrawing facial animation, mixing.
New materials are released over weeks, and avatar implementation in communications stops at two or three videos.
We explain how this is addressed: where the facial expression reference is fixed, how articulation quality is checked on control phrases, and what is needed for a digital avatar for a brand to be assembled from ready-made blocks rather than from scratch for each story.
Avatar capabilities: release speed and delivery quality
The video production cycle used to hit the shooting day: location, lighting, takes, editing — two to three weeks for release, and any text edit meant a new shift. With an avatar, material is assembled from a script: speech synthesis, facial animation, rendering. An edit takes one regeneration, not calling the crew to the set.
Without a shooting day, the main volume of content is produced — two to three dozen videos per month, including vertical and horizontal versions of one script.
Changing language or tone means swapping the audio track, not a new recording, so versions for different markets do not double the release schedule.
Delivery consistency no longer depends on a person’s schedule.
Pose, speech tempo, position in frame, and background are taken from the scene template and repeat frame by frame: in a series of dozens of releases, there is no “today the speaker looks different.” The speaker’s image is fixed once and does not change because of vacation, illness, or a busy schedule.
Under the hood are precomputed animation fragments and a library of standard phrases: repeating blocks are not calculated anew. Rendering runs on Unreal Engine and professional accelerators, so a one-minute piece is assembled in minutes.
The result is that the content plan is no longer tied to shoot dates, and release frequency grows without a proportional increase in team size.
Comparison of modes: live dialogue, recording, and broadcasting
The three avatar operating modes differ not in the picture but in the source of the line.
In an interactive avatar, the answer is assembled on the spot, for a specific viewer question; in pre-recorded scenes, the text is fixed before publication; broadcasting with an operator leaves control to a human.
Hence different requirements for channel, latency, and failure class: in some cases reaction speed is critical, in others — a verified wording that will not diverge from the brand’s regulations.
| Mode | When it fits | Limitations |
|---|---|---|
| Interactive dialogue | Consultations, Q&A, stands and support where each viewer comes with their own request | Latency for speech recognition; outside the described topic, the avatar falls back to a general answer and loses accuracy |
| Pre-recorded scenes | Greeting, training, regular announcements — wherever verified delivery matters | A question outside the script remains unanswered; any edit means re-recording the scene |
| Broadcasting with an operator | Presentations, conferences, streams with reaction to the audience | A person is needed to manage the course of the speech; higher channel load, possible desynchronization of gestures and speech |
In practice, the modes are combined: scenes cover typical requests, interactivity — non-standard questions, broadcasting — public speeches.
We review the limitations of the modes at the design stage and fix them in the description of the avatar’s behavior — what it does and does not do, so that the viewer does not receive promises the script cannot keep.
Work stages: from speaker image to going live
A digital avatar for a brand is not a separate video but a managed system: image, voice, and behavior are brought into a single loop that the client’s team then runs itself.
The order of stages is strict — each subsequent one relies on data from the previous one, and reordering means re-shooting or re-voicing.
- Image preparation. A test shoot of the speaker: articulation, facial expressions, angles, even lighting. We select takes where consonants are read unambiguously and discard the rest.
- Voice assembly. We record reference phrases with the required intonations and separately pronounce terms and brand names — otherwise the model smooths complex words into an even stream.
- Scenario scheme. We break the speech into lines and branches: what the avatar says itself, where it answers from the knowledge base, where it passes the question to a live specialist.
- Behavior setup. Speech tempo, pauses for meaning, facial expressions on key phrases, correct reaction to interruption and clarification.
- Testing on live questions. A run on real audience formulations, including inconvenient and compound ones. Based on the results, we adjust the knowledge base and confidence thresholds.
- Handover of control. Access, scenario update regulations, instructions for the editor — the team changes lines and materials without our involvement.
Duration is determined by the number of iterations in testing, not by the amount of shooting: the run on live questions takes the most time.
Regulations, image rights, and the material library
An avatar remains a brand asset only when the team runs it without us. Therefore we hand over not finished videos but a working kit: image description, usage rules, scenario templates, and the change procedure — in the client’s repository, under their version control.
- Image rights — written consent of the person or an agreement with the rights holder, recording channels, term, and territories of use, as well as prohibited contexts. The document sits next to the master model, not in a chat.
- Avatar usage regulations — permissible wording and tone, the set of emotions and speech tempo, restrictions on statements on behalf of the brand about products, conditions, and timelines.
- Template library — base scenes: greeting, product introduction, answers to typical questions. A unified background, lighting, and clothing so that a new scene can be shot in one pass without reassembling the image.
- Scenario schemes — the structure of the video with notes on where lip sync is needed and where voice-over and cutaways are sufficient.
- Team instructions — a step-by-step document: how to prepare text, which edits break articulation, how to assemble a video, and where to put the result.
- Update procedure — versioning of the master model and voice: a new version does not displace the old one until it has passed acceptance on a test video.
- Change log — who, when, and why the image was edited. If versions diverge, there is always a point of return to the working configuration.
Project: a speaker without studio shoots and shift schedules
What it was before
The speaker was a department head, and his calendar was the main constraint of the project. A series of training materials of twelve modules required twelve shooting days, and a window for recording was found once every three to four weeks.
Any edit of wording after methodological proofreading meant reassembling the crew and re-shooting the block entirely: the series release cycle stretched to two months, and some modules were already outdated by the time of publication.
What was changed and what the client received
We assembled a digital avatar of the speaker from a setup recording: we fixed angles, lighting, and manner of speech, and approved the base video image and the rules for its use in materials.
Then the module is assembled from the script — the text is edited in the script, assembling a forty-minute fragment takes hours rather than a shooting day, and the whole series does not need to be re-shot.
Series release was reduced from eight weeks to six working days, shooting days — one setup instead of twelve.
Re-shoots when updating regulations — zero versus six in the previous cycle, and typical answers to recurring questions are covered by short inserts without a new recording.
Akademiya «Osnova» formulates the result as follows: “Previously we adjusted to the speaker’s schedule and postponed updates until the next recording.
Now we make an edit in the script and a day later see a finished module — while the speaker remains recognizable and speaks in his own manner.”
How live dialogue differs from a pre-assembled video
The boundary is who sets the timing of the line.
In an interactive answer, the avatar receives a question, recognizes speech, selects a reaction, and synthesizes voice with facial animation right in the conversation — the delay between question and answer becomes the main quality metric.
In batch recording, the same frames are computed in advance and once, and the viewer gets a file that plays without computation on their side. Platforms of the HeyGen class can do both modes, but their implementation requirements diverge.
Hence different scenario preparation. For live dialogue, an intent map is written: several formulations of the same question, behavior rules for interruption and pauses, a separate “I don’t know” branch — without it the avatar confidently answers off-topic.
For a video, the script is closer to an edit sheet: line length and transitions to graphics matter, and variability only hurts.
| Mode | When it fits | Limitations |
|---|---|---|
| Interactive answer | stand, webinar, answers to unplanned questions | answer latency, a knowledge base and its regular revision are needed |
| Batch recording | training, presentations, advertising with fixed text | does not answer new questions, editing text means reassembly |
Choosing a mode is primarily choosing the amount of support. An interactive loop lives as long as it is maintained: the knowledge base becomes outdated, wording and product facts change, and without revision the avatar starts answering by old rules.
A recorded video after acceptance requires almost no attention, but changing one number means a new assembly.
Technical concerns: quality, updates, edits
Questions before signing on a digital avatar for a brand are usually not about style but about behavior in inconvenient situations: close-up, unplanned line, outdated text in frame. Below are direct answers; we will address the rest based on your source materials.
How realistic is the face in close-up?
We build the model from face scans and check it in close-ups under harsh side light and in profile — without blur that masks defects.
What will the avatar answer to an unplanned question?
It answers from your knowledge base or passes the dialogue to an operator; the threshold for switching to a live person is configurable. We mark synthetic video according to AI Act.
How are lines and appearance updated?
We change text and voice without re-shooting — articulation is recalculated for the new line. Changing the image is a separate iteration.
How is the edit procedure organized?
We collect notes in one list and close them in a round, so the number of rounds does not grow. First facial expressions and close-ups, then cosmetics.
What technical limitations are there?
Limits: monologue duration, set of languages and angles, stability during a sharp head turn. We state them at the start and propose a workaround — an edit cut or a live speaker.
Let’s discuss the task: we’ll assemble the speaker image and come back with a plan
To make the project estimate substantive rather than “by eye,” it is enough to prepare a few things for the discussion. Nothing needs to be assembled specially — materials the team already has will do, and we will record and close the gaps during the preparation stage.
- Examples of appearance and delivery: photos, videos, recordings of speeches — from them we describe the speaker’s image and range of facial expressions.
- Voice: whether there is a recording of the live voice, whether synthetic voice-over is needed, and in which languages the avatar should speak.
- Usage scenarios: where the avatar appears — website, training modules, presentations — and how long one line lasts.
- Text drafts: speech theses or frequent questions the speaker answers.
- Limitations: brand requirements, rights to source materials, who approves the final image.
- Guidelines on timelines and the format of the result: where it is embedded and in what form it is expected.
After the conversation, we come back with a project estimate, a solution outline, and a stage plan — with approval points and the scope of work at each step.
If some materials are not ready yet, this does not slow down the start of work: we will break down the collection by stages and suggest what is critical and what can be clarified along the way.