Voice Tuning & Customization - Burki Voice AI Docs

🎛️ Voice Tuning Mastery

Fine-tune voice characteristics across all TTS providers to create the perfect voice experience for your application. Learn stability, similarity, style, and provider-specific controls.

Overview of Voice Controls

🎯 Universal Voice Parameters

While each provider has unique features, these core concepts apply across most TTS services.

🎚️ Stability

Voice ConsistencyControls how consistent the voice sounds across different sentencesAvailable: ElevenLabs

🎯 Similarity

Voice AccuracyHow closely the output matches the original voice characteristicsAvailable: ElevenLabs

🎭 Style/Expression

Speaking StyleEmotional expression and speaking style variationAvailable: ElevenLabs, Inworld

Provider-Specific Controls

ElevenLabs
Inworld.ai
Deepgram & Resemble

🎭 ElevenLabs Voice Controls

The most comprehensive voice tuning options available.

Stability (0.0 - 1.0)

Controls voice consistency across sentences

Stability Settings Guide

Range	Effect	Best For	Example
0.0-0.2	Very expressive, inconsistent	Creative content, storytelling	Audiobooks with character voices
0.3-0.4	Expressive with some variation	Marketing content, presentations	Sales pitches, educational content
0.5-0.6	✅ Balanced (Recommended)	Business applications	Customer service, professional calls
0.7-0.8	Very consistent, less expressive	Technical content, instructions	Help desk, documentation
0.9-1.0	Extremely consistent, monotone	Announcements, alerts	System notifications, alerts

{
  "stability": 0.5,  // Recommended starting point
  "use_case": "balanced_professional"
}

Similarity Boost (0.0 - 1.0)

Controls how accurately the voice matches the original

Similarity Settings Guide

Range	Effect	Best For	Trade-off
0.0-0.4	Creative interpretation	Unique voice variations	Less like original voice
0.5-0.7	Balanced accuracy	Most applications	Good balance of creativity/accuracy
0.75	✅ Optimal (Recommended)	Production use	Best overall quality
0.8-0.9	Very accurate	Brand consistency	May sound slightly robotic
0.95-1.0	Extremely accurate	Voice cloning	Potential quality degradation

{
  "similarity_boost": 0.75,  // Sweet spot for most uses
  "note": "Recommended by ElevenLabs"
}

Style (0.0 - 1.0)

Controls speaking style and expressiveness

Style Settings Guide

Range	Effect	Best For	Personality
0.0	✅ Natural baseline	Business calls	Professional, neutral
0.1-0.3	Slight style variation	Customer service	Friendly, approachable
0.4-0.6	Moderate expression	Marketing content	Engaging, enthusiastic
0.7-0.9	High expressiveness	Entertainment	Dramatic, animated
1.0	Maximum style variation	Character voices	Highly expressive, theatrical

{
  "style": 0.0,  // Keep at 0.0 for business applications
  "business_rule": "Higher values can sound unprofessional"
}

Speaker Boost

Enhanced audio quality and clarity

✅ Enabled (Recommended)

Benefits:

Clearer voice quality
Reduced background noise
Better phone call clarity
Enhanced speech intelligibility

Best for: All applications

❌ Disabled

When to use:

Specific audio pipeline requirements
Custom post-processing needs
Legacy system compatibility

Trade-off: Lower audio quality

Latency Optimization (0-3)

Setting	Latency	Quality	Best For
0	~50ms	Lower	Experimental ultra-low latency
1	✅ Fastest vendor profile	Good	✅ Phone calls (Recommended)
2	Balanced vendor profile	Better	General applications
3	~250ms	Best	High-quality content creation

🎭 Inworld Emotional Controls

Emotional markup and contextual expression controls.

Emotional Markup

Direct emotional control through text tags

Basic Emotions

[happy] Great to hear from you!
[sad] I'm sorry about that issue.
[excited] This is fantastic news!
[concerned] Let me help you with that.
[confident] I can definitely assist you.
[grateful] Thank you for your patience.

Best Practices:

Use 1-2 emotion tags per sentence maximum
Match emotions to content context
Test different emotions with your chosen voice
Don’t overuse - natural text is still important

Speaking Styles

[whispering] This is confidential information.
[emphasizing] This is VERY important.
[casual] Hey there, how's it going?
[professional] Thank you for contacting our office.
[friendly] I'm happy to help you today!

Usage Guidelines:

Use styles that match your brand voice
Consider your audience and context
Test styles with different voices
Combine with basic emotions carefully

Language-Specific Tuning

Language	Voice Recommendations	Emotional Range	Production Ready
English	Ashley, Alex, Aria	Full range	✅ Production
Spanish	Diego, Lupita	Good range	✅ Production
French	Hélène, Mathieu	Good range	✅ Production
German	Johanna, Josef	Limited range	🟡 Beta
Chinese	Yichen, Xiaoyin	Limited range	🟡 Beta
Other	Various	Basic range	🟡 Beta

Model Selection for Expression

inworld-tts-1

Balanced Expression

Good emotional range
Consistent quality
✅ Recommended for business
~200ms latency

inworld-tts-1.5-max

Maximum Expression

Enhanced emotional range
More dramatic expressions
Best for creative content
~250ms latency

⚡ Speed-Focused Providers

These providers focus on speed and custom voices rather than extensive tuning options.

Deepgram Aura

Limited but effective controls:

Voice Selection Strategy

Choose the right voice for your use case:

Voice	Personality	Best For
Asteria	Warm, professional	Customer service
Luna	Soft, gentle	Healthcare, support
Stella	Clear, bright	Announcements
Orion	Authoritative	Business calls
Helios	Energetic	Marketing
Thalia	Professional	Corporate

Audio Format Optimization

Phone calls:

{
  "encoding": "mulaw",
  "sample_rate": 8000,
  "voice": "aura-asteria-en"
}

High quality:

{
  "encoding": "linear16",
  "sample_rate": 24000,
  "voice": "aura-2-thalia-en"
}

Resemble AI

Custom voice optimization:

Voice Training Quality

Optimize your custom voice training:

Training Data: 3-10 minutes of clear speech
Consistency: Same speaker, environment, microphone
Content: Diverse sentences and emotions
Quality: Clean audio, minimal background noise

# Voice training optimization
training_guidelines = {
    "duration": "3-10 minutes",
    "sample_rate": "22kHz or higher",
    "format": "WAV (preferred) or MP3",
    "content": "diverse_sentences_with_emotions",
    "environment": "quiet_professional_setting"
}

Streaming Configuration

WebSocket streaming optimization:

{
  "sample_rate": 8000,
  "precision": "MULAW",
  "output_format": "wav",
  "streaming": true
}

Best for phone calls and real-time applications

Recommended Settings by Use Case

Phone Calls
Customer Service
Content Creation
Multilingual Apps

📞 Phone Call Optimization

Settings optimized for clear, professional phone conversations.

ElevenLabs Phone Setup

{
  "model": "eleven_flash_v2_5",
  "voice": "rachel",
  "stability": 0.5,
  "similarity_boost": 0.75,
  "style": 0.0,
  "use_speaker_boost": true,
  "latency": 1
}

Deepgram Phone Setup

{
  "model": "aura-2-asteria-en",
  "encoding": "mulaw",
  "sample_rate": 8000
}

Inworld Phone Setup

{
  "model": "inworld-tts-1",
  "voice": "Ashley",
  "language": "en",
  "text": "[professional] Thank you for calling. [helpful] How may I assist you?"
}

Key Principles:

Prioritize clarity over expressiveness
Use phone-compatible audio formats
Keep emotional variation moderate
Enable speaker boost when available

🎧 Customer Service Excellence

Warm, helpful, and professional voice settings.

ElevenLabs Customer Service

{
  "voice": "bella",
  "stability": 0.6,
  "similarity_boost": 0.75,
  "style": 0.1,
  "use_speaker_boost": true
}

Inworld Customer Service

[friendly] Hello! [helpful] I'm here to assist you today. 
[understanding] I completely understand your concern. 
[confident] I can definitely help you resolve this.

Voice Selection:

ElevenLabs: Bella (warm), Rachel (professional)
Deepgram: Asteria (warm), Luna (gentle)
Inworld: Ashley (friendly), Aria (professional)
Resemble: Custom friendly customer service voice

Key Principles:

Warm and approachable tone
Consistent but not monotone
Slight emotional variation for empathy
Clear pronunciation for understanding

🎬 Content & Media

Engaging, expressive settings for videos, podcasts, and presentations.

ElevenLabs Content Creation

{
  "model": "eleven_v3",
  "voice": "antoni",
  "stability": 0.4,
  "similarity_boost": 0.7,
  "style": 0.3,
  "use_speaker_boost": true,
  "latency": 2
}

Inworld Content Creation

[excited] Welcome to our latest episode! 
[curious] Have you ever wondered about this topic? 
[explaining] Let me break this down for you.

Voice Selection:

ElevenLabs: Antoni (authoritative), Bella (engaging)
Inworld: Hades (dramatic), Aria (professional)
Resemble: Custom brand spokesperson voice

Key Principles:

Higher expressiveness acceptable
Quality over speed
Match voice to content personality
Use emotional variation for engagement

🌍 Global Applications

Settings optimized for multiple languages and international audiences.

ElevenLabs Multilingual

{
  "model": "eleven_v3",
  "language": "es",
  "stability": 0.5,
  "similarity_boost": 0.75,
  "style": 0.0
}

Inworld Multilingual

{
  "voice": "Diego",
  "language": "es",
  "text": "[amigable] ¡Hola! [servicial] ¿Cómo puedo ayudarte?"
}

Language-Specific Recommendations:

Language	Provider	Voice	Notes
Spanish	ElevenLabs/Inworld	Rachel/Diego	Native pronunciation
French	ElevenLabs/Inworld	Bella/Hélène	Natural accent
German	ElevenLabs/Inworld	Antoni/Johanna	Clear pronunciation
Chinese	Inworld	Yichen/Xiaoyin	Mandarin optimized
English	All providers	Various	Best quality available

Key Principles:

Use native voices when available
Test pronunciation with native speakers
Consider cultural communication styles
Adjust formality levels appropriately

Voice Testing & Optimization

🧪 Systematic Voice Testing

Develop a systematic approach to test and optimize your voice settings.

Testing Framework

Baseline Testing

Test with provider default settings using your actual content

Parameter Sweeping

Systematically adjust one parameter at a time

A/B Testing

Compare different settings with real users or stakeholders

Production Monitoring

Monitor voice quality and user feedback in live applications

Iterative Improvement

Continuously refine based on real-world usage data

Testing Script Examples

def test_elevenlabs_settings():
    test_cases = [
        {
            "name": "Conservative Business",
            "settings": {
                "stability": 0.6,
                "similarity_boost": 0.75,
                "style": 0.0,
                "use_speaker_boost": True
            }
        },
        {
            "name": "Balanced Professional", 
            "settings": {
                "stability": 0.5,
                "similarity_boost": 0.75,
                "style": 0.1,
                "use_speaker_boost": True
            }
        },
        {
            "name": "Expressive Friendly",
            "settings": {
                "stability": 0.4,
                "similarity_boost": 0.7,
                "style": 0.2,
                "use_speaker_boost": True
            }
        }
    ]
    
    test_text = "Hello! Thank you for calling our customer service line. How may I assist you today?"
    
    for test_case in test_cases:
        print(f"Testing: {test_case['name']}")
        # Generate audio with settings
        # Collect feedback or metrics

def test_inworld_emotions():
    emotion_tests = [
        "[professional] Thank you for calling our support line.",
        "[friendly] Hi there! How can I help you today?",
        "[helpful] I'd be happy to assist you with that.",
        "[understanding] I completely understand your concern.",
        "[confident] I can definitely resolve this for you."
    ]
    
    voices = ["Ashley", "Alex", "Aria"]
    
    for voice in voices:
        for emotion_text in emotion_tests:
            print(f"Testing {voice}: {emotion_text}")
            # Generate and evaluate audio

def compare_providers():
    test_text = "Welcome to our service. How may I help you today?"
    
    provider_configs = {
        "elevenlabs": {
            "voice": "rachel",
            "stability": 0.5,
            "similarity_boost": 0.75
        },
        "deepgram": {
            "voice": "aura-asteria-en"
        },
        "inworld": {
            "voice": "Ashley",
            "text": "[professional] " + test_text
        }
    }
    
    # Test each provider and collect metrics
    for provider, config in provider_configs.items():
        print(f"Testing {provider}")
        # Generate audio and measure:
        # - Latency
        # - Audio quality
        # - User preference

Common Tuning Mistakes

⚠️ Avoid These Pitfalls

Learn from common voice tuning mistakes to save time and improve results.

Over-Optimization

Problem: Adjusting too many parameters at onceSolution:

Change one parameter at a time
Test each change thoroughly
Keep notes on what works
Use A/B testing for comparisons

Example: Don’t change stability, similarity, and style simultaneously

Extreme Settings

Problem: Using values at the far ends of ranges (0.0 or 1.0)Solution:

Start with recommended ranges
Use extreme values only for specific effects
Test thoroughly before production use
Consider user experience impact

Example: style: 1.0 often sounds unnatural for business use

Ignoring Use Case

Problem: Using the same settings for different applicationsSolution:

Create setting profiles for different use cases
Consider your audience and context
Test with actual content types
Adjust based on user feedback

Example: Phone call settings ≠ podcast settings

Neglecting Voice Selection

Problem: Focusing only on parameters, ignoring voice choiceSolution:

Voice selection is often more important than fine-tuning
Test multiple voices with your content
Consider voice personality match
Use provider recommendations

Example: Wrong voice + perfect settings < Right voice + default settings

Advanced Optimization Techniques

Dynamic Settings
Content-Aware Tuning
User Preference Learning

Adjust settings based on context or content type

class DynamicVoiceSettings:
    def __init__(self):
        self.settings_profiles = {
            "greeting": {
                "stability": 0.6,
                "style": 0.1,
                "emotion": "[friendly]"
            },
            "problem_solving": {
                "stability": 0.5,
                "style": 0.0,
                "emotion": "[helpful]"
            },
            "closing": {
                "stability": 0.5,
                "style": 0.1,
                "emotion": "[grateful]"
            }
        }
        
    def get_settings(self, context):
        return self.settings_profiles.get(context, self.settings_profiles["greeting"])

Adjust based on content analysis

def analyze_and_tune(text):
    # Analyze content characteristics
    word_count = len(text.split())
    has_questions = '?' in text
    has_exclamations = '!' in text
    formality_score = calculate_formality(text)
    
    # Adjust settings based on analysis
    if formality_score > 0.8:
        return {"stability": 0.6, "style": 0.0}
    elif has_exclamations:
        return {"stability": 0.4, "style": 0.2}
    elif has_questions:
        return {"stability": 0.5, "style": 0.1}
    else:
        return {"stability": 0.5, "style": 0.0}

Learn from user feedback and behavior

class VoicePreferenceLearner:
    def __init__(self):
        self.feedback_data = []
        self.current_settings = default_settings()
        
    def record_feedback(self, settings, rating, context):
        self.feedback_data.append({
            "settings": settings,
            "rating": rating,
            "context": context,
            "timestamp": time.now()
        })
        
    def optimize_settings(self):
        # Analyze feedback patterns
        # Adjust settings based on highest-rated combinations
        # Return optimized settings
        pass

📚 Provider Guides

Detailed Provider Information:

ElevenLabs Setup - Premium quality controls
Deepgram Configuration - Speed optimization
Inworld Emotions - Emotional markup
Resemble Custom Voices - Brand voice creation

🛠️ Advanced Topics

Next Steps:

Troubleshooting Guide - Fix common issues
Best Practices - Production optimization
AI Configuration - System-wide settings

🎯 Perfect Your Voice Settings

Use this guide to systematically optimize your TTS voice settings. Start with recommended defaults, test systematically, and refine based on your specific use case and user feedback.

🎛️ Voice Tuning Mastery

​Overview of Voice Controls

🎯 Universal Voice Parameters

🎚️ Stability

🎯 Similarity

🎭 Style/Expression

​Provider-Specific Controls

🎭 ElevenLabs Voice Controls

​Stability (0.0 - 1.0)

​Similarity Boost (0.0 - 1.0)

​Style (0.0 - 1.0)

​Speaker Boost

✅ Enabled (Recommended)

❌ Disabled

​Latency Optimization (0-3)

🎭 Inworld Emotional Controls

​Emotional Markup

​Language-Specific Tuning

​Model Selection for Expression

inworld-tts-1

inworld-tts-1.5-max

⚡ Speed-Focused Providers

​Deepgram Aura

​Resemble AI

​Recommended Settings by Use Case

📞 Phone Call Optimization

​ElevenLabs Phone Setup

​Deepgram Phone Setup

​Inworld Phone Setup

🎧 Customer Service Excellence

​ElevenLabs Customer Service

​Inworld Customer Service

🎬 Content & Media

​ElevenLabs Content Creation

​Inworld Content Creation

🌍 Global Applications

​ElevenLabs Multilingual

​Inworld Multilingual

​Voice Testing & Optimization

🧪 Systematic Voice Testing

​Testing Framework

​Testing Script Examples

​Common Tuning Mistakes

⚠️ Avoid These Pitfalls

​Advanced Optimization Techniques

​Related Guides

📚 Provider Guides

🛠️ Advanced Topics

🎯 Perfect Your Voice Settings

Overview of Voice Controls

Provider-Specific Controls

Stability (0.0 - 1.0)

Similarity Boost (0.0 - 1.0)

Style (0.0 - 1.0)

Speaker Boost

Latency Optimization (0-3)

Emotional Markup

Language-Specific Tuning

Model Selection for Expression

Deepgram Aura

Resemble AI

Recommended Settings by Use Case

ElevenLabs Phone Setup

Deepgram Phone Setup

Inworld Phone Setup

ElevenLabs Customer Service

Inworld Customer Service

ElevenLabs Content Creation

Inworld Content Creation

ElevenLabs Multilingual

Inworld Multilingual

Voice Testing & Optimization

Testing Framework

Testing Script Examples

Common Tuning Mistakes

Advanced Optimization Techniques

Related Guides