voice-design

AI 语音设计

使用 ElevenLabs、Qwen3-TTS 和其他平台，为内容挑选并创建理想的 AI 语音，使其语音特征与品牌个性和受众相匹配。

适用场景

• 为视频旁白挑选 AI 语音
• 在不同内容中打造一致的品牌声音
• 克隆语音以实现规模化制作
• 比较语音合成平台
• 根据描述设计语音特征
• 为不同角色或用途挑选多种语音

方法论基础

来源：ElevenLabs + Qwen3-TTS + 语音设计最佳实践

核心原则：“语音占视频效果的 50%”——语音挑选或生成不当，会破坏沉浸感。最好的 AI 语音，是听众察觉不到它由 AI 生成的语音。要做到这一点，就需要让语音特征（年龄、性别、音调、语速）符合内容类型和受众预期。

重要性：对于大多数使用场景，AI 语音合成已达到人类水平的品质，让大规模内容创作成为可能。但技术效果取决于语音选择。无论语音听起来多么自然，不匹配的语音都会削弱内容效果。

Claude 的职责与由你决定的事项

Claude 的职责	由你决定
规划制作工作流	最终创意方向
建议技术方案	设备和工具选择
创建模板和检查清单	质量标准
总结最佳实践	品牌和语音决策
生成脚本大纲	脚本最终审核

此技能的功能

1. 让语音与品牌相匹配 - 将品牌属性转化为语音特征
2. 选择最佳平台 - 根据需求在 ElevenLabs、Qwen3-TTS 和其他方案中进行选择
3. 根据描述设计语音 - 根据文本提示词创建自定义语音
4. 管理语音一致性 - 在不同项目中保持语音一致
5. 为多语音项目选角 - 为对话或角色挑选相辅相成的语音

使用方法

选择 AI 语音

Help me choose an AI voice for [content type].
Brand: [personality]
Audience: [who]
Content: [describe]
Platform preference: [if any]

设计自定义语音

Design a voice for my brand:
Brand personality: [traits]
Target audience: [who]
Use cases: [where voice will be used]

比较平台

Compare voice platforms for my needs:
Volume: [how much content]
Languages: [which]
Budget: [range]
Features needed: [cloning, real-time, etc.]

说明

设计 AI 语音时，请遵循以下方法：

步骤 1：明确语音需求

选择平台或语音之前，先记录你的需求。

## Voice Requirements Worksheet

### Brand Alignment
**Brand personality** (3-5 traits):
**Tone of voice** (formal/casual/playful/etc.):
**Existing brand sounds** (if any):

### Audience Match
**Primary audience** (age, context):
**What voice would they trust?**
**What voice would feel authentic to them?**

### Technical Requirements
**Languages needed**:
**Monthly volume** (minutes/hours):
**Real-time needed?** (yes/no):
**Voice cloning needed?** (yes/no):
**Budget** (monthly):

### Content Types
□ Long-form narration (courses, audiobooks)
□ Short-form video (social, ads)
□ Conversational (chatbots, assistants)
□ Character voices (multiple speakers)
□ Localization (same voice, multiple languages)

---

步骤 2：选择平台

根据需求选择平台。

## Platform Decision Matrix (2026)

### ElevenLabs
**Best for**: Premium quality, voice cloning, multilingual
**Pricing**: $5-330/mo
**Languages**: 29+
**Voice cloning**: Yes (from $22/mo)
**Latency**: 75ms (Flash v2.5)

**Choose if**:
- Quality is top priority
- Need professional voice cloning
- Require many languages with same voice
- Budget allows $20+/mo

### Qwen3-TTS (Open Source)
**Best for**: Self-hosted, zero marginal cost, privacy
**Pricing**: Free (+ GPU costs)
**Languages**: 10
**Voice cloning**: Yes (zero-shot from 3 seconds)
**Latency**: 97ms streaming

**Choose if**:
- Processing sensitive data locally
- High volume (cost per minute matters)
- Technical capability to self-host
- Need real-time streaming

### Murf.ai
**Best for**: Professional voiceover, video workflow
**Pricing**: $19-99/mo
**Languages**: 45+
**Voice cloning**: Limited
**Special**: "Say It My Way" intonation control

**Choose if**:
- Need studio voiceover quality
- Video production workflow
- Team collaboration needed
- Want precise pronunciation control

### OpenAI TTS
**Best for**: Simple integration, developer-focused
**Pricing**: $15/M characters
**Languages**: Limited
**Voices**: 6 presets (alloy, echo, fable, onyx, nova, shimmer)

**Choose if**:
- Already using OpenAI ecosystem
- Simple API integration needed
- Don't need customization
- Light usage

### Budget Decision

| Budget | Recommendation |
|--------|----------------|
| $0 | Qwen3-TTS (self-hosted) |
| $5-20/mo | ElevenLabs Starter |
| $20-50/mo | ElevenLabs Creator (with cloning) |
| $100+/mo | ElevenLabs Pro or Murf Pro |
| High volume | Self-hosted Qwen3-TTS |

---

步骤 3：将品牌特征转化为语音特征

将品牌属性转化为语音参数。

## Brand-to-Voice Translation

### Voice Attributes

| Brand Trait | Voice Translation |
|-------------|-------------------|
| Professional | Lower pitch, measured pace, clear articulation |
| Friendly | Mid-pitch, warm tone, slight smile quality |
| Authoritative | Deep, resonant, slower pace, confident pauses |
| Energetic | Higher pitch variation, faster pace, dynamic range |
| Trustworthy | Steady, consistent, neutral accent, clear |
| Innovative | Modern quality, subtle processing, distinctive |
| Warm | Rich mid-tones, soft consonants, unhurried |
| Premium | Controlled, polished, slight reverb/space |

### Voice Parameter Guide

**Pitch Range**:
- Low: Authority, seriousness, gravitas
- Mid: Versatility, approachability
- High: Energy, youth, friendliness

**Pace**:
- Slow: Premium, thoughtful, serious content
- Medium: Most content, versatile
- Fast: Energetic, urgent, young audience

**Accent**:
- Neutral: Universal appeal, no specific region
- Regional: Authenticity for specific markets
- International: European/British for sophistication (to US ears)

---

步骤 4：通过描述设计语音（ElevenLabs）

ElevenLabs 支持通过文本描述设计语音。

## Voice Design Prompts

### Template
"A [gender] voice in their [age range], with a [accent] accent.
The voice is [tone qualities] with [delivery characteristics].
[Additional characteristics or limitations]."

### Examples

**Corporate Explainer**:
"A male voice in his late 30s, with a neutral American accent.
The voice is warm and professional with clear articulation and
measured pacing. Sounds like a trusted advisor, not a salesman."

**E-learning Instructor**:
"A female voice in her early 40s, with a slight British accent.
The voice is encouraging and patient with a natural, conversational
delivery. Sounds like a supportive teacher who makes complex
topics accessible."

**Tech Product Demo**:
"A young male voice in his late 20s, with a West Coast American accent.
The voice is confident and energetic with a modern, casual delivery.
Sounds knowledgeable but not condescending, like explaining to a
friend who's also into tech."

**Luxury Brand**:
"A female voice in her 30s, with a subtle French accent.
The voice is sophisticated and understated with elegant pacing
and restrained emotion. Sounds exclusive but welcoming, never rushed."

### Tips for Better Results
- Be specific about age (not just "young" but "late 20s")
- Describe the feeling, not just mechanics
- Reference the context/listener relationship
- Iterate: try 3-5 variations, pick the best

---

步骤 5：多语音选角

当内容需要多位说话者时。

## Voice Casting for Multi-Speaker Content

### Dialogue Principles

**Contrast**:
- Different pitches (one higher, one lower)
- Different timbres (one warm, one bright)
- Different energies (one measured, one dynamic)

**Cohesion**:
- Similar quality level
- Compatible accents
- Both feel "from the same world"

### Example Cast

**Corporate Training Video (3 voices)**:

| Role | Voice Type | Platform Choice |
|------|-----------|-----------------|
| Narrator | Authoritative female, 40s | ElevenLabs "Charlotte" |
| Employee A | Friendly male, 30s | ElevenLabs "Daniel" |
| Employee B | Energetic female, 20s | ElevenLabs "Elli" |

**Podcast-Style Explainer (2 voices)**:

| Role | Voice Type | Characteristics |
|------|-----------|-----------------|
| Host | Warm male, mid-30s | Conversational, asks questions |
| Expert | Authoritative female, 40s | Knowledgeable, explains |

### Casting Checklist

□ Voices are clearly distinguishable by ear
□ Voices complement (not clash)
□ Power dynamic appropriate for content
□ All voices pass the "would I trust this person?" test
□ Consistent quality/processing across all voices

---

步骤 6：管理语音一致性

在不同项目中保持语音一致。

## Voice Consistency System

### Documentation

Create a Voice ID Card for each brand voice:

```markdown
## Voice ID: [Brand Name] Primary

**Platform**: ElevenLabs
**Voice ID/Name**: [voice identifier]
**Created**: [date]

### Specifications
- Base voice: [name or description]
- Stability: 0.7 (or setting used)
- Clarity: 0.8 (or setting used)
- Style exaggeration: 0 (or setting used)

### Usage Guidelines
- Primary use: [main content types]
- Never use for: [inappropriate contexts]
- Pair with: [complementary voices]

### Reference Sample
[Link to audio sample]
[Link to key content using this voice]

### Settings History
- v1.0 (date): Initial settings
- v1.1 (date): Adjusted clarity for better consonants

质量控制

□ 使用参考样本对照检查新内容 □ 如果使用克隆语音，检查是否出现偏移 □ 确认处理方式和效果保持一致 □ 在多种播放设备上进行测试 □ 定期请外部人员试听检查

## Examples

### Example 1: SaaS Product Demo Voice

**Context**: Choosing an AI voice for product demo videos.

**Input**:
> Help me choose an AI voice for our SaaS product demos. Brand is professional but friendly, audience is B2B decision makers (30-50 years old), and we produce 4-5 videos per month.

**Output**:

## Voice Selection: SaaS Demo Videos

### Requirements Analysis
- **Brand**: Professional + friendly (approachable authority)
- **Audience**: B2B decision makers, 30-50
- **Volume**: 4-5 videos/month (~30-60 min total)
- **Implicit needs**: Trustworthy, clear, not salesy

### Platform Recommendation: ElevenLabs Creator ($22/mo)

**Why**:
- Quality matches professional B2B expectations
- Volume fits within plan limits
- Professional voice library
- Room to clone founder voice later if desired

### Voice Selection

**Primary Voice**: Rachel (ElevenLabs)
- Female, American, mid-30s sound
- Professional warmth, clear articulation
- Works for both feature demos and testimonial reads
- Doesn't sound "AI" to most ears

**Alternative**: Josh (ElevenLabs)
- Male, American, early 40s sound
- Authority without stuffiness
- Good for technical deep-dives

### Settings

Voice: Rachel Stability: 0.75 (natural variation) Clarity + Similarity: 0.80 (prioritize clarity) Style: 0 (neutral, professional)

### Sample Script Test

Before committing, test with representative content:

"DataFlow integrates with your existing tools in minutes,
not months. Here's how: Simply connect your Salesforce instance
using our secure OAuth flow, and we'll automatically map your
custom fields. No IT tickets required."

→ Listen for: Natural flow, clear tech terms, appropriate pace

---

### Example 2: Multi-Language Brand Voice

**Context**: Expanding content to multiple markets with consistent voice.

**Input**:
> We need to create content in English, French, and Spanish with a consistent brand voice. Same "person" across all languages. Budget is flexible for the right solution.

**Output**:

## Multi-Language Voice Strategy

### Challenge
Maintain consistent brand "character" across languages—not just translation, but the same perceived person speaking.

### Recommended Platform: ElevenLabs Pro ($99/mo)

**Why ElevenLabs for multilingual**:
- Same voice can speak 29+ languages natively
- Preserves tone and character across languages
- Better than separate native voices for brand consistency
- Dubbing feature maintains lip-sync

### Implementation Approach

**Option 1: Design Voice, Use Across Languages**

1. Create perfect English voice using Voice Design
2. Use that exact voice for French and Spanish
3. ElevenLabs handles accent authentically per language

**Voice Design Prompt**:
"A warm, confident voice in the early 30s. Gender-neutral leaning
slightly feminine. Clear, professional articulation with a modern,
international quality. Should sound equally at home in New York,
Paris, or Madrid. Approachable expert energy."

**Option 2: Clone Founder/Spokesperson**

If you have a real person who embodies the brand:
1. Clone their voice (30+ min sample)
2. Use clone for all languages
3. Their "essence" transfers, accent adapts

### Language-Specific Notes

| Language | Consideration |
|----------|--------------|
| English | Base voice, primary development |
| French | Slightly slower pace, French pronunciation patterns |
| Spanish | Choose Castilian vs Latin American variant |

### Quality Control

- Native speaker review for each language
- Check for unnatural pronunciation of brand terms
- Verify numbers and dates sound correct
- Test technical vocabulary

## Checklists & Templates

### Voice Selection Checklist

选择之前

□ 记录品牌个性 □ 明确受众 □ 列出内容类型 □ 估算内容量 □ 确认预算 □ 明确所需语言

选择流程

□ 筛选出 3-5 个候选语音 □ 使用实际脚本内容进行测试 □ 在目标设备上试听（手机、笔记本电脑） □ 收集团队反馈 □ 测试听觉疲劳情况（连续聆听 5 分钟以上） □ 确认不同样本内容中的语音保持一致

选择之后

□ 记录语音设置 □ 保存参考样本 □ 制定使用指南 □ 使用制作内容进行测试 □ 必要时规划本地化

---

### Voice Platform Comparison

快速参考

需求	最佳选择
高级品质	ElevenLabs
零成本	Qwen3-TTS（可自托管）
语音克隆	ElevenLabs Creator+
29+ 种语言	ElevenLabs
视频工作流	Murf.ai
OpenAI 生态系统	OpenAI TTS
实时	Qwen3-TTS 或 ElevenLabs Flash
数据隐私	Qwen3-TTS（可自托管）

## 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 - Voice design and cloning guides
- Qwen3-TTS vs ElevenLabs Comparison - ByteIota
- Best Text-to-Speech AI 2026 - AIML API review
- Murf AI Review - Voice design workflows

## Related Skills

- [voice-localization](../voice-localization/) - Same voice across languages
- [voiceover-direction](../voiceover-direction/) - Working with human talent
- [sonic-branding](../sonic-branding/) - Brand audio identity
- [video-testimonial](../video-testimonial/) - Customer video content

---

## Skill Metadata (Internal Use)

```yaml
name: voice-design
category: audio
subcategory: voice
version: 1.0
author: MKTG Skills
source_expert: ElevenLabs, Qwen3-TTS
source_work: Platform Documentation, Industry Comparisons
difficulty: intermediate
estimated_value: $200-1,000 per voice design project
tags: [ai-voice, tts, elevenlabs, voice-synthesis, brand-voice]
created: 2026-01-26
updated: 2026-01-26