5.3Animation, lip sync & voice
The essentials in 30 seconds
- The three-brick chain: portrait animation (img2vid) + voice synthesis + lip sync — each brick tested separately.
- Non-negotiable rules: never the face or voice of a real person without explicit documented consent; transparency about the generated nature is mandatory.
- Short and validated: short sequences (lip sync degrades over duration), never binding statements without validation.
Running case — Sonia creates an ethical virtual presenter
Sonia's client wants an explainer video “presented” with no shoot. Sonia, a freelance graphic designer in Algiers, suggests a fully AI-generated character (not a client's face), a short text validated by the client, a synthetic voice adjusted in tone and pace, then lip sync on a short sequence — with a clear “AI-generated content” disclosure. Sequel to case 5.2 — Sonia makes a character speak without ever touching a real person's face.
Sequel to case 5.2 — Sonia assembles img2vid, voice and lip sync with written consent and a transparency disclosure.
1. The three-brick chain
Making a character speak combines: 1) image animation (img2vid — micro-movements, blinking, slight head moves that bring a portrait to life), 2) voice synthesis (a voice generated from text, with control over tone, pace and language), 3) lip sync (matching the mouth to the audio track). Sonia tests each brick separately before assembling.
Technically, lip sync stays imperfect on long, close-up shots — short sequences work far better. It is a constraint to build in from the writing of the text.
2. Legitimate uses
Legitimate uses are many: educational content (a virtual presenter explaining), internal videos and training, translation and dubbing into other languages, avatars for product information — all at a cost far below a shoot. Sonia sees in it a way to deliver an explainer video to the client with no studio.
Including the most legitimate case: cloning your own voice or face, since you consent by definition — translations, corrections without reshooting, learning modules. But even there, secure access to the clone and stay transparent with the audience.
3. Non-negotiable rules
This is the ground with the highest ethical risks in the whole guide. Never the voice or face of a real person without their explicit, documented consent: that is the very definition of a deepfake, with serious legal and human consequences (section 0.4). Transparency: clearly indicate that it is generated content — an undisclosed synthetic presenter deceives the audience.
Caution about content: never have an avatar say binding statements (medical, financial, legal advice) without validation and disclosure. And keep the mirror reflex: faced with a dubious video, question the source, look for independent confirmation, beware the emotional context that short-circuits verification.
🔎 Concrete example — Sonia's ethical spoken sequence
Generated character, short text, clear disclosure.
GENERATED character (or person with written consent).
Short text → voice (synthesis) → animation + lip sync.
→ clear “AI-generated content” disclosure.Reading: each brick is tested separately before assembly in the edit.
Lesson: short sequence — lip sync degrades over duration.
Section vocabulary
- img2vid
- Animating a still portrait to give it natural micro-movements.
Simple example: Sonia animates the generated portrait: blinking, slight head moves.
What it is not: a complete film generated in one block.
Reflex: does my portrait look alive even before speaking?
- Voice synthesis
- Generating a voice from text, with control over tone, pace and language.
Simple example: Sonia generates the voice of the validated text, adjusting tone and pace.
What it is not: a studio recording with an actor.
Reflex: does the tone fit the client's message?
- Lip sync
- Synchronising mouth movements with the audio track.
Simple example: Sonia synchronises the character's mouth to the generated voice.
What it is not: perfect synchronisation guaranteed on long shots.
Reflex: is my sequence short enough to hold?
- Avatar
- A synthetic character presenting educational or product content.
Simple example: Sonia's virtual presenter explains the client's offer.
What it is not: a real person filmed without knowing.
Reflex: is my character generated or consented in writing?
- Documented consent
- Explicit, recorded consent before any use of someone's face or voice.
Simple example: Sonia has a written consent signed, limited to a defined use.
What it is not: a vague or assumed verbal consent.
Reflex: can I prove the consent in writing?
In practice — Create a short spoken sequence (ethically)
- Use a GENERATED character (not a real person) or a person with written consent.
- Write a short text, then generate the voice with a speech-synthesis service, adjusting tone and pace.
- Animate the portrait and apply lip sync; favour a SHORT sequence (synchronisation degrades over duration).
- Add a clear disclosure that the content is AI-generated, then integrate the sequence into your edit.
Ready-to-say script — proposing an ethical avatar to the client
“I can create a virtual presenter for your explainer video with no shoot: a fully AI-generated character, a short text you validate, an adjusted synthetic voice, a short sequence for good synchronisation. I will add a clear disclosure that the content is AI-generated — and I will never use a real person's face or voice without their written consent.”
To say to the client to lock the ethical framework before any avatar production.
Copy-paste template — ethical spoken-sequence sheet
ETHICAL SPOKEN-SEQUENCE SHEET
CHARACTER: generated / real with written consent of ___ (use: ___)
TEXT: short, validated by ___ on ___ · binding statements? YES / NO
VOICE: synthesis, tone ___ · pace ___ · language ___
LIP SYNC: sequence ___ s (short) · tested YES / NO
DISCLOSURE: “AI-generated content” placed ___Solo / small budget
Sonia alone: one generated character and a thirty-second text are enough for a first spoken sequence.
Small business / small team
Text reviewed by two people before voice generation: you never have an avatar say improvisation.
SME in Algiers
In Algiers, Sonia has the client validate the text and the transparency disclosure before assembly.
Starting team
Pair: one produces the sequence, the other checks written consent, disclosure and absence of binding statements.
✓ Best practices
- Use a generated character or a consenting person (written consent, defined use).
- Clearly disclose the AI nature of the content, even for your own clone.
- Favour short sequences and have any binding statements validated.
✗ Common pitfalls
- Using a real person's face or voice without consent: it is a deepfake.
- Releasing an avatar with no transparency disclosure: the audience is deceived.
- Having it say binding statements (health, finance, law) without validation.
Key takeaways
- Typical chain: portrait animation (img2vid) + voice synthesis + lip sync.
- Legitimate uses: education, internal training, dubbing/translation, product information — at very low cost.
- NON-NEGOTIABLE RULES: never a real person's face or voice without explicit documented consent; transparency mandatory.
- Lip sync stays imperfect on long, close-up shots: favour short sequences.
48h challenge — apply tomorrow
Fill in the spoken-sequence sheet (template above) for 1 real short text: do the written consent and the AI disclosure appear in it?
- I can describe the img2vid + voice + lip sync chain in one sentence.
- My character is generated or consented in writing.
- My short sequence carries a transparency disclosure.
Frequently asked questions
Can I clone my own voice or face to create my content?
Yes — a legitimate use since you consent (translations, corrections, modules). Precautions: secure access to the clone, check the deletion terms for your data, stay transparent with your audience, and validate any binding statements as if you were really speaking. For others: explicit, written consent limited to a defined use.
How do I recognise a deepfake, and how do I protect myself?
Visual clues (face/neck transitions, teeth, eyes, off lip sync) disappear as progress comes — the real reflexes are contextual: question the source, look for independent confirmation, beware emotional urgency, and require validation via a second channel against targeted scams. For protection: rely on traceability (C2PA, section 5.5) and limit the circulation of long sequences of your voice and face.
What is the non-negotiable rule for making an avatar speak?