Design and generate AI voices for your videos using ElevenLabs or Qwen3-TTS, with voice cloning, design by description, and lip-sync synchronization.
When to Use This Skill
- Create a brand voice - Define the vocal tone for a campaign
- Clone an existing voice - Reproduce a voice with authorization
- Design an original voice - Create a voice from a description
- Multi-character - Manage multiple voices in the same video
- AI video lip-sync - Synchronize voice and lip movements
- Localization - Adapt a voice into multiple languages
Methodology Foundation
Source: ElevenLabs Documentation + Qwen3-TTS (Alibaba) + PJ Ace workflow
Core Principle: "Voice is 50% of a video's impact. ElevenLabs offers premium quality, Qwen3-TTS offers open-source flexibility. The choice depends on budget and desired control."
Why This Matters: A poorly chosen or poorly generated voice breaks the illusion of an AI video. The right voice/character matching and professional audio quality make the difference between "AI slop" and credible content.
What Claude Does vs What You Decide
| Claude Does | You Decide |
|---|---|
| Structures production workflow | Final creative direction |
| Suggests technical approaches | Equipment and tool choices |
| Creates templates and checklists | Quality standards |
| Identifies best practices | Brand/voice decisions |
| Generates script outlines | Final script approval |
What This Skill Does
- Compares TTS solutions - ElevenLabs vs Qwen3-TTS
- Guides voice cloning - Process and requirements
- Structures voice design - Textual description of voices
- Manages multi-voice - Assignment and consistency
- Prepares lip-sync - Integration with Kling/Veo
How to Use
Design a voice for an ad
J'ai besoin d'une voix pour ma pub [produit]. Le ton doit être [description]. Aide-moi à la designer.
Clone a voice
Je veux cloner cette voix [description/échantillon]. Guide-moi à travers le process avec [ElevenLabs/Qwen3-TTS].
Manage multiple characters
J'ai 3 personnages dans ma vidéo: [descriptions]. Crée le casting vocal.
Instructions
Step 1: Choose the TTS solution
## Comparatif ElevenLabs vs Qwen3-TTS
| Critère | ElevenLabs | Qwen3-TTS |
|---------|------------|-----------|
| **Prix** | $5-$330/mois | Gratuit (open-source) |
| **Qualité** | Excellente | Excellente |
| **Voice cloning** | 1-5 min audio | 3 sec audio |
| **Langues** | 29+ | 10 (EN, FR, DE, ES, IT, PT, RU, ZH, JP, KO) |
| **Latence** | ~200ms | 97ms (streaming) |
| **Self-hosted** | Non | Oui |
| **Voice design** | VoiceLab | Description textuelle |
| **API** | Cloud | Local ou cloud |
### Arbre de décision
**Q1: Budget disponible?**
- $0 (gratuit) → Qwen3-TTS
- $5-$100/mois → ElevenLabs Starter/Creator
- $100+/mois → ElevenLabs Pro
**Q2: Besoin de self-hosting?**
- Oui (données sensibles) → Qwen3-TTS
- Non → ElevenLabs ou Qwen3-TTS
**Q3: Langue requise parmi FR, EN, DE, ES, IT, PT, RU, ZH, JP, KO?**
- Oui → Les deux fonctionnent
- Non (autre langue) → ElevenLabs
**Q4: Latence critique (real-time)?**
- Oui → Qwen3-TTS (97ms)
- Non → Les deux
Step 2: Voice Design by description
## Template de Description de Voix
### Caractéristiques de base
**Genre:** [ ] Masculin [ ] Féminin [ ] Non-binaire
**Âge apparent:** [20s / 30s / 40s / 50s / 60s+]
**Registre:** [ ] Grave [ ] Medium [ ] Aigu
### Qualités vocales
**Texture:**
[ ] Lisse/Veloutée [ ] Rauque/Gravelly
[ ] Nasale [ ] Claire [ ] Résonante
**Énergie:**
[ ] Calme/Posée [ ] Dynamique [ ] Intense
[ ] Chaleureuse [ ] Froide/Distante
**Rythme:**
[ ] Lent/Délibéré [ ] Modéré [ ] Rapide
[ ] Varié (storytelling)
### Accent/Origine
**Accent:** [Ex: Français neutre, British RP, Southern US, etc.]
**Particularités:** [Ex: légèrement rauque le matin, sourire dans la voix]
### Contexte d'utilisation
**Marque/Produit:** ________________________________
**Ton de la campagne:** ________________________________
**Personnage (si fiction):** ________________________________
### Exemples de référence (optionnel)
**Voix similaire à:** [Célébrité, personnage, pub connue]
**Éviter:** [Ce qu'on ne veut pas]
Example of a complete description:
## Voice Brief: Pub NeuroBoost
**Profil:**
Homme, 40s, registre grave-medium
**Qualités:**
- Texture résonante et autoritaire mais pas intimidante
- Énergie posée, confiante, légèrement inspirante
- Rythme délibéré avec pauses stratégiques
**Accent:**
Français international (pas d'accent régional marqué)
Diction parfaite, articulation claire
**Références:**
- Similaire à: Morgan Freeman mais version française
- Éviter: Ton commercial agressif, urgence artificielle
**Notes:**
Cette voix doit incarner l'expertise et la confiance.
Le spectateur doit sentir qu'il reçoit un conseil d'un
mentor plutôt qu'un pitch de vendeur.
Step 3: Voice cloning
ElevenLabs
## Process de clonage ElevenLabs
### Instant Clone (1-5 min audio)
**Qualité:** Bonne (80% fidélité)
**Usage:** Tests, itération rapide
1. Préparer audio source:
- 1-5 minutes de parole claire
- Pas de musique de fond
- Qualité minimum: 128kbps
- Formats: MP3, WAV, M4A
2. Dans ElevenLabs:
- Voice Lab → Add Voice → Instant Clone
- Upload audio
- Nommer la voix
- Tester avec phrase sample
### Professional Clone (30+ min audio)
**Qualité:** Excellente (95%+ fidélité)
**Usage:** Production commerciale
1. Préparer corpus audio:
- 30-60 minutes idéalement
- Variété d'émotions et tons
- Phrases complètes, pas de mots isolés
- Studio quality (256kbps+, pas de bruit)
2. Soumettre pour training
- Délai: 24-48h
- Coût: Inclus dans plan Pro+
### Paramètres de génération
- **Stability:** 50-70% (naturel) / 80%+ (consistant)
- **Clarity:** 75%+ recommandé
- **Style:** 0-100% selon expressivité souhaitée
Qwen3-TTS
## Process de clonage Qwen3-TTS
### Zero-shot Clone (3 sec audio)
**Qualité:** Très bonne
**Usage:** Toute production
1. Préparer référence:
- 3-10 secondes de parole claire
- Pas de bruit de fond
- Émotion neutre ou représentative
2. API Python:
```python
from qwen3_tts import Qwen3TTS
tts = Qwen3TTS()
# Cloner depuis référence
voice = tts.clone_voice(
reference_audio="reference.wav",
voice_name="my_voice"
)
# Générer avec la voix clonée
audio = tts.generate(
text="Texte à synthétiser",
voice=voice,
language="fr"
)
audio.save("output.wav")
Voice Design by text
# Créer une voix sans référence audio
voice = tts.design_voice(
description="A warm, confident male voice in his 40s, \
with a slight French accent, speaking slowly and \
deliberately with gravitas."
)
audio = tts.generate(
text="Votre texte ici",
voice=voice
)
Advanced parameters
- emotion: "neutral", "happy", "sad", "angry", "surprise"
- speed: 0.5 (slow) to 2.0 (fast)
- pitch: -10 to +10 (semitones)
---
### Step 4: Multi-personnages
Multi-character Vocal Casting
Casting template
| Character | Description | Voice | Source |
|---|---|---|---|
| [Name] | [Physical description/personality] | [Vocal specs] | [Clone/Design/Stock] |
Example: Ad with 3 characters
| Character | Description | Voice | Source |
|---|---|---|---|
| CEO Emma | 35 years old, confident, leader | Woman, medium, authoritative-warm | Design: "Confident female executive..." |
| Dev Tom | 28 years old, enthusiastic geek | Man, medium-high, fast | Stock: "Young professional male" |
| Client Marc | 50 years old, skeptical then convinced | Man, deep, hesitant→confident | Design: "Skeptical older businessman..." |
Differentiation rules
- Varied registers: Deep, Medium, High
- Different rhythms: Slow vs Fast
- Distinct accents: If appropriate to the context
- Contrasting energies: Calm vs Dynamic
Multi-voice workflow
- Generate each line separately
- Name files:
P1_Emma_Line01.wav - Assemble in audio timeline
- Check volume consistency (normalize to -6dB)
---
### Step 5: Lip-sync et intégration vidéo
Voice + AI Video Integration
Option A: Veo 3.1 (Native audio)
The voice is generated by Veo with the video.
Prompt incluant audio:
"[Scene description]. The character says:
'[Dialogue exact]'. Voice: confident male, 30s,
warm tone. Ambient: office sounds."
Limitation: Less control over the exact voice.
Option B: Kling 2.6 Motion Control
Lip-sync from "driving" video.
- Record the dialogue yourself
- Use Kling with driving video
- Lip-sync follows your performance
- Replace audio with AI voice in post
Workflow:
You → Record driving video with dialogue
Kling → Transfer lips to AI character
Post → Replace audio with ElevenLabs/Qwen voice
Option C: Separate generation + Post-sync
For maximum control.
- Generate video without audio (Runway, Pika)
- Generate voice separately
- Synchronize in post-production
- Adjust timing manually if needed
Sync tools:
- DaVinci Resolve (free)
- Adobe Premiere
- D-ID (automatic sync)
Lip-sync Checklist
- Voice timing matches lip movement
- Natural pauses aligned
- Volume balanced with ambience
- No visible delay (max 2-3 frames)
---
### Step 6: Export et spécifications audio
Final Audio Specifications
Output format
- Codec: AAC or WAV
- Sample rate: 48kHz (video standard)
- Bit depth: 24-bit (WAV) or 256kbps (AAC)
- Channels: Stereo or Mono depending on use
Recommended levels
- Voiceover: -6dB to -3dB peak
- Dialogue: -12dB to -6dB
- Ambiance: -18dB to -12dB
- Music: -12dB to -9dB (under voice)
Basic mastering
- Normalize all voice clips to -6dB
- Apply light compression (ratio 2:1)
- EQ: cut below 80Hz, slight boost 2-4kHz (clarity)
- Limit at -1dB to avoid clipping
Deliverables
📁 Audio_[Project]/
├── voiceover/
│ ├── VO_final_master.wav
│ └── VO_stems/
│ ├── P1_lines.wav
│ ├── P2_lines.wav
│ └── ...
├── sfx/
│ └── [sound effects]
├── music/
│ └── [background music]
└── mix/
└── final_mix.wav
## Examples
### Example 1: Voice design pour pub premium
**Context**: Marque de montres de luxe, pub cinématique 30s
**Input**:
> J'ai une pub pour une marque de montres haut de gamme. Style très premium, inspiré des pubs Omega ou Rolex. Je veux une voix qui incarne l'excellence et l'héritage.
**Output**:
## Voice Brief: Montres Heritage Collection
### Profil cible
Gender: Male Age: 55-65 years Register: Low, deep
### Caractéristiques détaillées
Texture: Resonant, velvety, with a slight patina (like noble wood or vintage leather)
Energy: Calm, contemplative, never rushed Every word is chosen, every pause is intentional
Rhythm: Very slow, deliberate 4-5 words per breath Pauses of 1-2 seconds between sentences
Diction: Perfect, aristocratic without being snobbish Precise consonants, rich vowels
### Prompt ElevenLabs (Voice Design)
Create a distinguished male voice in his 60s. Deep, resonant baritone with natural gravitas. Speaking style: slow, deliberate, contemplative. British Received Pronunciation with slight warmth. Texture like aged whiskey - smooth with depth. The voice of someone who has seen much and speaks only when it matters.
### Prompt Qwen3-TTS
```python
voice = tts.design_voice(
description="""
Distinguished older gentleman, 60s, deep baritone.
British accent, RP. Speaking very slowly and
deliberately, with natural authority and warmth.
Voice texture is rich and resonant, like aged oak.
Contemplative, never rushed. Each word carries weight.
Think: nature documentary narrator meets luxury brand.
"""
)
Script with directions
[PAUSE 2s]
"Time..." [PAUSE 1.5s]
"...is not measured." [PAUSE 1s]
"It is crafted." [PAUSE 2s]
"Heritage Collection." [PAUSE 1s]
"Since 1892."
Technical settings
- Stability: 85% (very consistent for luxury)
- Speed: 0.7x (slower)
- No music under the voice (voice only, then music)
Why this works: For luxury, less is more. A deep, slow, calm voice conveys confidence and timelessness. Pauses create tension and perceived value.
Example 2: Comedy multi-character
Context: Offbeat SaaS ad, sitcom style with 3 characters
Input:
I have an Office-style ad with 3 characters: the incompetent boss, the exasperated employee, and the overly enthusiastic intern. Create the voice casting.
Output:
Voice Casting: "Office Chaos" Ad
Character 1: Michel (The Boss)
Profil:
- Homme, 50 ans
- Confiant mais incompétent
- Parle trop fort et trop vite
- S'interrompt lui-même
Voix:
- Medium-aigu (pas grave malgré l'autorité voulue)
- Nasale légèrement
- Volume variable (starts loud, gets louder)
- Phrases inachevées
Prompt ElevenLabs:
"Middle-aged man who thinks he's charismatic but isn't.
Slightly nasal voice, medium-high pitch. Speaks too
loudly and too fast, often interrupting himself.
French accent, corporate buzzwords enthusiast.
Michael Scott energy."
Character 2: Sophie (The Exasperated One)
Profil:
- Femme, 35 ans
- Compétente, fatiguée
- Sarcasme subtil
- Soupirs audibles
Voix:
- Medium, légèrement grave pour femme
- Sèche, précise
- Monotone quand exaspérée
- Soupirs comme ponctuation
Prompt Qwen3-TTS:
"Professional woman in her 30s, tired of everything.
Dry, slightly deadpan delivery. French, neutral accent.
Subtle sarcasm in every line. Occasional audible sighs.
The straight man in every comedy duo."
Character 3: Théo (The Intern)
Profil:
- Homme, 22 ans
- Trop enthousiaste
- Voix qui monte en fin de phrase
- Acquiesce à tout
Voix:
- Medium-aigu
- Énergique, rapide
- Upspeak (fin de phrase montante)
- Ponctué de "super!", "génial!"
Prompt:
"Young man, early 20s, overly enthusiastic intern.
High-medium pitch, speaks quickly with upward
inflection at end of sentences. French, sounds
like he just discovered coffee. Every statement
sounds like an excited question."
Dialogue example
MICHEL: (loud) "Bon, l'équipe! J'ai une GRANDE nouvelle—
enfin, moyenne—non, grande!"
SOPHIE: (flat) "[soupir] ...C'est la réunion quotidienne."
THÉO: (excited) "Oh WOW! Une grande nouvelle? C'est GÉNIAL!"
MICHEL: "Théo comprend, LUI. Donc, on va—
comment ça s'appelle—pivoter!"
SOPHIE: "...On a pivoté hier."
THÉO: "RE-pivoter! J'ADORE re-pivoter!"
Production notes
- Generate each character separately
- Michel: boost 3-4kHz (more "present")
- Sophie: slight room reverb (emotional distance)
- Théo: compression to contain peaks
Why this works: The three voices are immediately distinct in register, rhythm and energy. The contrast creates the comedy - Michel's chaos, Sophie's calm, Théo's excess.
Checklists & Templates
Voice Design Checklist
## Validation Voice Design
### Brief complet
- [ ] Genre et âge définis
- [ ] Registre spécifié (grave/medium/aigu)
- [ ] Texture décrite (lisse/rauque/etc)
- [ ] Énergie et rythme indiqués
- [ ] Accent précisé
- [ ] Références incluses
### Génération
- [ ] Prompt testé avec phrase sample
- [ ] Qualité audio vérifiée (pas de glitches)
- [ ] Volume normalisé
- [ ] Cohérence avec brand voice
### Multi-personnages
- [ ] Voix suffisamment distinctes
- [ ] Registres variés
- [ ] Énergies contrastées
- [ ] Test d'écoute ensemble
Voice Brief Template
## Voice Brief: [Projet]
### Identité
**Projet/Marque:** ________________________________
**Type de contenu:** [ ] Pub [ ] Explainer [ ] Narration [ ] Dialogue
**Durée totale:** __________ secondes
### Profil vocal
**Genre:** [ ] M [ ] F [ ] Non-binaire
**Âge:** _______ ans
**Registre:** [ ] Grave [ ] Medium [ ] Aigu
### Caractéristiques
**Texture:** ________________________________
**Énergie:** ________________________________
**Rythme:** ________________________________
**Accent:** ________________________________
### Contexte émotionnel
**L'auditeur doit ressentir:** ________________________________
**Éviter:** ________________________________
### Référence
**Similaire à:** ________________________________
### Solution technique
[ ] ElevenLabs (budget: $_____/mois)
[ ] Qwen3-TTS (self-hosted)
[ ] Clone d'une voix existante
### Script
[Paste the script with pause directions]
Comparative costs
## Budget TTS
### ElevenLabs
| Plan | Prix/mois | Caractères | Équivalent |
|------|-----------|------------|------------|
| Free | $0 | 10k | ~2 min |
| Starter | $5 | 30k | ~6 min |
| Creator | $22 | 100k | ~20 min |
| Pro | $99 | 500k | ~100 min |
### Qwen3-TTS (Self-hosted)
| Composant | Coût |
|-----------|------|
| GPU (RTX 3090) | ~$800 one-time |
| Cloud GPU (A10) | ~$1/heure |
| Hébergement | Variable |
| Génération | Illimité |
### Recommandation par volume
- < 5 min/mois → ElevenLabs Free + Qwen3-TTS
- 5-20 min/mois → ElevenLabs Creator ($22)
- 20-100 min/mois → ElevenLabs Pro ($99)
- > 100 min/mois → Qwen3-TTS self-hosted
Skill Boundaries
What This Skill Does Well
- Structuring audio production workflows
- Providing technical guidance
- Creating quality checklists
- Suggesting creative approaches
What This Skill Cannot Do
- Replace audio engineering expertise
- Make subjective creative decisions
- Access or edit audio files directly
- Guarantee commercial success
References
- ElevenLabs Documentation
- ElevenLabs Voice Cloning
- Qwen3-TTS GitHub
- Qwen3-TTS Blog
- MKTG Skills - AI Video Ecosystem Study
Related Skills
- ai-video-prompting - Audio integration in Veo prompts
- ai-video-qa - Audio quality verification
- copywriting-ogilvy - Writing voice scripts
Skill Metadata
- Mode: cyborg
name: ai-voice-design
category: video
subcategory: production
version: 1.0
author: MKTG Skills
source_expert: ElevenLabs + Qwen3-TTS + PJ Ace
source_work: TTS Documentation
difficulty: intermediate
estimated_value: $500-2000 (voice design + production)
tags: [video, ai, voice, tts, elevenlabs, qwen, cloning, audio]
created: 2026-01-25
updated: 2026-01-25