Skip to main content
What you’ll learn: Detailed breakdown of latencies across all Vodex system components including ASR, AI models, and TTS providers with performance characteristics and optimization notes.
Understanding latency is crucial for optimizing your AI calling campaigns. This guide provides comprehensive latency information for all system components to help you make informed decisions about your configuration.

​
ASR (Automatic Speech Recognition) Latencies

Speech recognition is the first step in processing user input. Different ASR providers offer varying latency characteristics based on their processing approach.

​
Streaming ASRs

Real-time speech recognition that processes audio as it’s being spoken:
  • Alpha Echo V2
  • Nova Echo
Ultra-Low Latency Processing
MetricValue
Base Latency250ms
Processing TypeStreaming
Best ForReal-time conversations, interactive scenarios
Characteristics:
  • Fastest streaming ASR option
  • Optimized for conversational AI
  • Minimal delay in speech recognition
Streaming ASR Timing: For streaming ASRs, we implement adaptive timing - either 300ms OR 1 second depending on user activity to prevent cutting off users mid-sentence.

​
Non-Streaming ASRs

Batch processing ASRs that analyze complete audio segments:
Enterprise-Grade Recognition
MetricValue
Processing Latency2 seconds
VAD Time300ms (fixed)
Total Latency~2.3 seconds
Best for: High-accuracy requirements, complex audio processing
Advanced AI-Powered ASR
MetricValue
Processing Latency700ms
VAD Time300ms (fixed)
Total Latency~1 second
Best for: Multi-language support, challenging audio conditions
Optimized Whisper Implementation
MetricValue
Processing Latency400ms
VAD Time300ms (fixed)
Total Latency~700ms
Best for: Fast Whisper processing, quality audio recognition
Microsoft Cloud ASR
MetricValue
Processing Latency800ms
VAD Time300ms (fixed)
Total Latency~1.1 seconds
Best for: Enterprise integrations, Microsoft ecosystem
Specialized Voice Recognition
MetricValue
Processing Latency400ms
VAD Time300ms (fixed)
Total Latency~700ms
Best for: Specialized voice processing, custom implementations
Non-Streaming ASR Note: All non-streaming ASRs include a fixed Voice Activity Detection (VAD) time of 300ms to ensure complete speech capture before processing.

​
AI Model Latencies

Time to first token generation - the critical metric for conversational responsiveness.
Important: Listed latencies represent time to first token. Complete response generation typically adds +300ms from the first token delivery.
Latency vs Capability Trade-off: Lower latency models are typically smaller LLMs that cannot handle longer prompts, complex rules, or extensive context. Choose based on your specific use case requirements.

​
Vodex Optimized Models

Specially tuned models optimized for voice conversations:
  • Spark Series
  • Performance Comparison
Vodex Custom-Based Architecture
ModelFirst Token LatencyStatusBest For
Spark1.5s✅ StableComplex reasoning, detailed conversations
Spark Flash400ms✅ StableBalanced performance, standard interactions
Spark Flash Lite200ms✅ StableQuick responses, simple tasks
Performance Characteristics:
  • Optimized for voice conversations
  • Excellent reasoning capabilities
  • Consistent performance across scenarios

​
OpenAI Models

Industry-leading AI models with varying performance characteristics:
Advanced Reasoning Models
ModelFirst Token LatencyStatusCharacteristics
GPT-4o800ms✅ StableMultimodal capabilities, advanced reasoning
GPT-4o Mini500ms✅ StableEfficient multimodal processing
GPT-4.1 Mini500ms✅ StableEnhanced efficiency, cost-effective
Best for: Complex conversations, detailed analysis, multimodal interactions
Cutting-Edge AI Models
ModelFirst Token LatencyStatusCharacteristics
GPT-5 Nano350ms⚠️ Latency issuesUltra-efficient processing, limited context
GPT-5 Mini200ms⚠️ Latency issuesFast responses, simplified reasoning
GPT-5 Chat500ms✅ StableOptimized for conversations
GPT-5 Status: GPT-5 Nano and Mini are currently experiencing latency issues. Monitor performance closely if using these models.
Purpose-Built AI Solutions
ModelFirst Token LatencyStatusCharacteristics
OpenAI GSS1.2s✅ StableSpecialized processing, high accuracy
Best for: Specialized applications requiring high accuracy

​
Open Source & Alternative Models

High-performance alternatives to proprietary models:
  • Meta Llama
  • DeepSeek
Open Source Excellence
ModelFirst Token LatencyStatusCharacteristics
Llama 3.3 70B450ms✅ StableLarge parameter model, excellent reasoning
Llama 4 Maverick450ms✅ StableNext-generation open source
Advantages:
  • Open source flexibility
  • Privacy-focused processing
  • Customizable implementations
  • Cost-effective scaling

​
TTS (Text-to-Speech) Latencies

Time to first audio chunk - critical for maintaining conversation flow.
TTS Streaming: All TTS providers support streaming, allowing audio playback to begin as soon as the first chunk is available, reducing perceived latency.

​
ElevenLabs (Default Provider)

Premium AI voice synthesis with multiple performance tiers:
  • Turbo Series
  • Flash Series
  • Multilingual Series
High-Quality Voice Synthesis
ModelFirst Chunk LatencyQualityBest For
Turbo 2250msHighProfessional conversations, customer service
Turbo 2.5250msEnhancedPremium voice quality, sales calls
Characteristics:
  • Premium voice quality
  • Natural emotional expression
  • Multiple voice personalities
  • Excellent for professional use
ElevenLabs Selection: Use Flash series for real-time interactions where speed is critical, and Turbo series for professional scenarios where voice quality is paramount.

​
Alternative TTS Providers

Additional voice synthesis options for specialized needs:
Advanced Voice Technology
MetricValue
First Chunk Latency350ms
QualityHigh
AvailabilityContact support
Best for:
  • Specialized voice requirements
  • Custom voice training
  • Enterprise implementations
WaveNet Technology
MetricValue
First Chunk Latency400ms
QualityUltra-realistic
AvailabilityContact support
Best for:
  • Ultra-realistic voice requirements
  • Multi-language campaigns
  • Google Cloud integrations
Microsoft Neural Voices
MetricValue
First Chunk Latency800ms
QualityEnterprise-grade
AvailabilityContact support
Best for:
  • Enterprise applications
  • Microsoft ecosystem integration
  • Multi-language support

​
Latency Optimization Strategies

​
Configuration Recommendations

  • Real-Time Conversations
  • Balanced Performance
  • High-Quality Processing
Ultra-Low Latency SetupRecommended Configuration:
  • ASR: Alpha Echo V2 (250ms)
  • Model: Spark Flash Lite (200ms) or GPT-5 Mini (200ms)*
  • TTS: ElevenLabs Flash 2.5 (95ms)
Total Pipeline Latency: ~545ms
Limitations: Ultra-low latency models cannot handle complex prompts, extensive context, or sophisticated rules. Best for simple, straightforward interactions only.
*GPT-5 Mini currently experiencing latency issues

​
Performance Monitoring

1

Baseline Measurement

Establish Performance Baselines
  • Monitor end-to-end conversation latency
  • Track component-specific performance
  • Document peak and average response times
2

Optimization Testing

A/B Test Configurations
  • Compare different ASR providers
  • Test various AI model combinations
  • Evaluate TTS provider performance
3

Continuous Monitoring

Ongoing Performance Tracking
  • Set up latency alerts
  • Monitor for performance degradation
  • Track user experience metrics

​
Technical Considerations

​
Latency Factors

Network Latency

External Factors
  • Geographic distance to servers
  • Network congestion and routing
  • Internet service provider performance
  • CDN and edge server optimization

Processing Load

System Performance
  • Server capacity and utilization
  • Concurrent request handling
  • Model loading and initialization
  • Resource allocation efficiency

Audio Quality

Input Characteristics
  • Audio sample rate and quality
  • Background noise levels
  • Speaker clarity and volume
  • Connection stability

​
Best Practices

Speech Recognition Best Practices
  • Choose streaming ASR for real-time scenarios
  • Consider non-streaming for accuracy-critical applications
  • Account for VAD timing in latency calculations
  • Test with representative audio samples
AI Model Optimization
  • Match model complexity to use case requirements
  • Monitor for model-specific latency issues
  • Consider fallback options for unstable models
  • Balance reasoning capability with response speed
Voice Synthesis Optimization
  • Use streaming TTS for reduced perceived latency
  • Select voice quality appropriate for use case
  • Consider multiple TTS providers for redundancy
  • Test voice quality across different scenarios

​
Next Steps

Ready to optimize your latency? Use this comprehensive guide to select the optimal configuration for your specific use case and performance requirements.
Recommended Actions:
  1. Assess your requirements - Determine if you need real-time, balanced, or high-quality processing
  2. Test configurations - Experiment with different component combinations
  3. Monitor performance - Track latency metrics in production
  4. Optimize iteratively - Continuously refine based on real-world performance
Need help with configuration? Contact support@vodex.ai for personalized latency optimization assistance.
Configure these latency-related settings in your Vodex dashboard:

ASR Configuration

Speech Recognition SettingsConfigure ASR providers and settings in Call Settings and Advanced Settings.
  • Choose between streaming and non-streaming ASR
  • Configure language codes and detection
  • Set up custom ASR parameters

AI Model Selection

Model ConfigurationSelect and configure AI models in Call Settings.
  • Choose from Vodex optimized models
  • Configure OpenAI, Llama, and other providers
  • Balance latency vs capability requirements

Voice/TTS Settings

Text-to-Speech ConfigurationConfigure voice providers and settings in Call Settings.
  • Select ElevenLabs Turbo or Flash series
  • Configure alternative TTS providers
  • Optimize voice quality vs speed
⌘I