> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vodex.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# ⚡ Understanding Latency

> Comprehensive guide to understanding latencies across ASR, AI models, and TTS systems in Vodex

<Info>
  **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.
</Info>

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:

<Tabs>
  <Tab title="Alpha Echo V2">
    **Ultra-Low Latency Processing**

    | Metric              | Value                                          |
    | ------------------- | ---------------------------------------------- |
    | **Base Latency**    | 250ms                                          |
    | **Processing Type** | Streaming                                      |
    | **Best For**        | Real-time conversations, interactive scenarios |

    **Characteristics:**

    * Fastest streaming ASR option
    * Optimized for conversational AI
    * Minimal delay in speech recognition
  </Tab>

  <Tab title="Nova Echo">
    **Balanced Streaming Performance**

    | Metric              | Value                                         |
    | ------------------- | --------------------------------------------- |
    | **Base Latency**    | 300ms                                         |
    | **Processing Type** | Streaming                                     |
    | **Best For**        | Standard voice interactions, customer service |

    **Characteristics:**

    * Reliable streaming performance
    * Good balance of speed and accuracy
    * Suitable for most use cases
  </Tab>
</Tabs>

<Note>
  **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.
</Note>

### Non-Streaming ASRs

Batch processing ASRs that analyze complete audio segments:

<AccordionGroup>
  <Accordion title="Google ASR">
    **Enterprise-Grade Recognition**

    | Metric                 | Value         |
    | ---------------------- | ------------- |
    | **Processing Latency** | 2 seconds     |
    | **VAD Time**           | 300ms (fixed) |
    | **Total Latency**      | \~2.3 seconds |

    **Best for:** High-accuracy requirements, complex audio processing
  </Accordion>

  <Accordion title="OpenAI Whisper">
    **Advanced AI-Powered ASR**

    | Metric                 | Value         |
    | ---------------------- | ------------- |
    | **Processing Latency** | 700ms         |
    | **VAD Time**           | 300ms (fixed) |
    | **Total Latency**      | \~1 second    |

    **Best for:** Multi-language support, challenging audio conditions
  </Accordion>

  <Accordion title="ElevenLabs Whisper">
    **Optimized Whisper Implementation**

    | Metric                 | Value         |
    | ---------------------- | ------------- |
    | **Processing Latency** | 400ms         |
    | **VAD Time**           | 300ms (fixed) |
    | **Total Latency**      | \~700ms       |

    **Best for:** Fast Whisper processing, quality audio recognition
  </Accordion>

  <Accordion title="Azure Speech Services">
    **Microsoft Cloud ASR**

    | Metric                 | Value         |
    | ---------------------- | ------------- |
    | **Processing Latency** | 800ms         |
    | **VAD Time**           | 300ms (fixed) |
    | **Total Latency**      | \~1.1 seconds |

    **Best for:** Enterprise integrations, Microsoft ecosystem
  </Accordion>

  <Accordion title="Genesis Echo">
    **Specialized Voice Recognition**

    | Metric                 | Value         |
    | ---------------------- | ------------- |
    | **Processing Latency** | 400ms         |
    | **VAD Time**           | 300ms (fixed) |
    | **Total Latency**      | \~700ms       |

    **Best for:** Specialized voice processing, custom implementations
  </Accordion>
</AccordionGroup>

<Warning>
  **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.
</Warning>

***

## AI Model Latencies

Time to first token generation - the critical metric for conversational responsiveness.

<Info>
  **Important:** Listed latencies represent time to first token. Complete response generation typically adds +300ms from the first token delivery.
</Info>

<Warning>
  **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.
</Warning>

### Vodex Optimized Models

Specially tuned models optimized for voice conversations:

<Tabs>
  <Tab title="Spark Series">
    **Vodex Custom-Based Architecture**

    | Model                | First Token Latency | Status   | Best For                                    |
    | -------------------- | ------------------- | -------- | ------------------------------------------- |
    | **Spark**            | 1.5s                | ✅ Stable | Complex reasoning, detailed conversations   |
    | **Spark Flash**      | 400ms               | ✅ Stable | Balanced performance, standard interactions |
    | **Spark Flash Lite** | 200ms               | ✅ Stable | Quick responses, simple tasks               |

    **Performance Characteristics:**

    * Optimized for voice conversations
    * Excellent reasoning capabilities
    * Consistent performance across scenarios
  </Tab>

  <Tab title="Performance Comparison">
    **Spark Model Selection Guide**

    <Card title="Spark" icon="brain">
      **Best for complex scenarios**

      * Sales conversations requiring detailed reasoning
      * Multi-step problem solving
      * Technical support discussions
      * **Trade-off:** Higher latency for better reasoning
    </Card>

    <Card title="Spark Flash" icon="bolt">
      **Balanced performance**

      * Customer service interactions
      * Standard business conversations
      * Appointment scheduling
      * **Sweet spot:** Good balance of speed and capability
    </Card>

    <Card title="Spark Flash Lite" icon="zap">
      **Ultra-fast responses**

      * Simple confirmations
      * Quick surveys
      * Basic information gathering
      * **Advantage:** Minimal latency for simple tasks
      * **Limitation:** Cannot handle complex prompts or extensive context
    </Card>
  </Tab>
</Tabs>

### OpenAI Models

Industry-leading AI models with varying performance characteristics:

<AccordionGroup>
  <Accordion title="GPT-4 Series">
    **Advanced Reasoning Models**

    | Model            | First Token Latency | Status   | Characteristics                             |
    | ---------------- | ------------------- | -------- | ------------------------------------------- |
    | **GPT-4o**       | 800ms               | ✅ Stable | Multimodal capabilities, advanced reasoning |
    | **GPT-4o Mini**  | 500ms               | ✅ Stable | Efficient multimodal processing             |
    | **GPT-4.1 Mini** | 500ms               | ✅ Stable | Enhanced efficiency, cost-effective         |

    **Best for:** Complex conversations, detailed analysis, multimodal interactions
  </Accordion>

  <Accordion title="GPT-5 Series (Next Generation)">
    **Cutting-Edge AI Models**

    | Model          | First Token Latency | Status            | Characteristics                             |
    | -------------- | ------------------- | ----------------- | ------------------------------------------- |
    | **GPT-5 Nano** | 350ms               | ⚠️ Latency issues | Ultra-efficient processing, limited context |
    | **GPT-5 Mini** | 200ms               | ⚠️ Latency issues | Fast responses, simplified reasoning        |
    | **GPT-5 Chat** | 500ms               | ✅ Stable          | Optimized for conversations                 |

    <Warning>
      **GPT-5 Status:** GPT-5 Nano and Mini are currently experiencing latency issues. Monitor performance closely if using these models.
    </Warning>
  </Accordion>

  <Accordion title="Specialized Models">
    **Purpose-Built AI Solutions**

    | Model          | First Token Latency | Status   | Characteristics                       |
    | -------------- | ------------------- | -------- | ------------------------------------- |
    | **OpenAI GSS** | 1.2s                | ✅ Stable | Specialized processing, high accuracy |

    **Best for:** Specialized applications requiring high accuracy
  </Accordion>
</AccordionGroup>

### Open Source & Alternative Models

High-performance alternatives to proprietary models:

<Tabs>
  <Tab title="Meta Llama">
    **Open Source Excellence**

    | Model                | First Token Latency | Status   | Characteristics                            |
    | -------------------- | ------------------- | -------- | ------------------------------------------ |
    | **Llama 3.3 70B**    | 450ms               | ✅ Stable | Large parameter model, excellent reasoning |
    | **Llama 4 Maverick** | 450ms               | ✅ Stable | Next-generation open source                |

    **Advantages:**

    * Open source flexibility
    * Privacy-focused processing
    * Customizable implementations
    * Cost-effective scaling
  </Tab>

  <Tab title="DeepSeek">
    **High-Performance Alternative**

    | Model             | First Token Latency | Status   | Characteristics                          |
    | ----------------- | ------------------- | -------- | ---------------------------------------- |
    | **Deepseek v3.1** | 300ms               | ✅ Stable | Fast processing, competitive performance |

    **Benefits:**

    * Excellent speed-to-performance ratio
    * Reliable processing
    * Cost-effective option
  </Tab>
</Tabs>

***

## TTS (Text-to-Speech) Latencies

Time to first audio chunk - critical for maintaining conversation flow.

<Info>
  **TTS Streaming:** All TTS providers support streaming, allowing audio playback to begin as soon as the first chunk is available, reducing perceived latency.
</Info>

### ElevenLabs (Default Provider)

Premium AI voice synthesis with multiple performance tiers:

<Tabs>
  <Tab title="Turbo Series">
    **High-Quality Voice Synthesis**

    | Model         | First Chunk Latency | Quality  | Best For                                     |
    | ------------- | ------------------- | -------- | -------------------------------------------- |
    | **Turbo 2**   | 250ms               | High     | Professional conversations, customer service |
    | **Turbo 2.5** | 250ms               | Enhanced | Premium voice quality, sales calls           |

    **Characteristics:**

    * Premium voice quality
    * Natural emotional expression
    * Multiple voice personalities
    * Excellent for professional use
  </Tab>

  <Tab title="Flash Series">
    **Ultra-Fast Voice Generation**

    | Model         | First Chunk Latency | Quality  | Best For                                       |
    | ------------- | ------------------- | -------- | ---------------------------------------------- |
    | **Flash 2**   | 95ms                | Good     | Real-time conversations, interactive scenarios |
    | **Flash 2.5** | 95ms                | Enhanced | Fast interactions with improved quality        |

    **Advantages:**

    * Ultra-low latency
    * Real-time conversation capability
    * Optimized for speed
    * Maintains good voice quality
  </Tab>

  <Tab title="Multilingual Series">
    **Advanced Multi-Language Support**

    | Model               | First Chunk Latency | Quality | Best For                                  |
    | ------------------- | ------------------- | ------- | ----------------------------------------- |
    | **Multilingual V2** | 1.2s                | High    | Multi-language campaigns, global outreach |

    **Characteristics:**

    * Advanced multilingual capabilities
    * High-quality voice synthesis across languages
    * Optimized for international campaigns
    * Supports diverse language requirements

    **Trade-off:**

    * Higher latency for comprehensive language support
    * Best when language quality is more important than speed
  </Tab>
</Tabs>

<Tip>
  **ElevenLabs Selection:** Use Flash series for real-time interactions where speed is critical, and Turbo series for professional scenarios where voice quality is paramount.
</Tip>

### Alternative TTS Providers

Additional voice synthesis options for specialized needs:

<AccordionGroup>
  <Accordion title="RimeLabs">
    **Advanced Voice Technology**

    | Metric                  | Value           |
    | ----------------------- | --------------- |
    | **First Chunk Latency** | 350ms           |
    | **Quality**             | High            |
    | **Availability**        | Contact support |

    **Best for:**

    * Specialized voice requirements
    * Custom voice training
    * Enterprise implementations
  </Accordion>

  <Accordion title="Google UR Realistic">
    **WaveNet Technology**

    | Metric                  | Value           |
    | ----------------------- | --------------- |
    | **First Chunk Latency** | 400ms           |
    | **Quality**             | Ultra-realistic |
    | **Availability**        | Contact support |

    **Best for:**

    * Ultra-realistic voice requirements
    * Multi-language campaigns
    * Google Cloud integrations
  </Accordion>

  <Accordion title="Azure Cognitive Services">
    **Microsoft Neural Voices**

    | Metric                  | Value            |
    | ----------------------- | ---------------- |
    | **First Chunk Latency** | 800ms            |
    | **Quality**             | Enterprise-grade |
    | **Availability**        | Contact support  |

    **Best for:**

    * Enterprise applications
    * Microsoft ecosystem integration
    * Multi-language support
  </Accordion>
</AccordionGroup>

***

## Latency Optimization Strategies

### Configuration Recommendations

<Tabs>
  <Tab title="Real-Time Conversations">
    **Ultra-Low Latency Setup**

    **Recommended 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

    <Warning>
      **Limitations:** Ultra-low latency models cannot handle complex prompts, extensive context, or sophisticated rules. Best for simple, straightforward interactions only.
    </Warning>

    <Warning>
      \*GPT-5 Mini currently experiencing latency issues
    </Warning>
  </Tab>

  <Tab title="Balanced Performance">
    **Speed-Quality Balance**

    **Recommended Configuration:**

    * **ASR:** Nova Echo (300ms)
    * **Model:** Spark Flash (400ms) or GPT-4o Mini (500ms)
    * **TTS:** ElevenLabs Turbo 2.5 (250ms)

    **Total Pipeline Latency:** \~950ms - 1050ms
  </Tab>

  <Tab title="High-Quality Processing">
    **Premium Quality Setup**

    **Recommended Configuration:**

    * **ASR:** OpenAI Whisper (700ms + 300ms VAD)
    * **Model:** Spark (1.5s) or GPT-4o (800ms)
    * **TTS:** ElevenLabs Turbo 2.5 (250ms)

    **Total Pipeline Latency:** \~1.8s - 2.25s
  </Tab>
</Tabs>

### Performance Monitoring

<Steps>
  <Step title="Baseline Measurement">
    **Establish Performance Baselines**

    * Monitor end-to-end conversation latency
    * Track component-specific performance
    * Document peak and average response times
  </Step>

  <Step title="Optimization Testing">
    **A/B Test Configurations**

    * Compare different ASR providers
    * Test various AI model combinations
    * Evaluate TTS provider performance
  </Step>

  <Step title="Continuous Monitoring">
    **Ongoing Performance Tracking**

    * Set up latency alerts
    * Monitor for performance degradation
    * Track user experience metrics
  </Step>
</Steps>

***

## Technical Considerations

### Latency Factors

<Card title="Network Latency" icon="network-wired">
  **External Factors**

  * Geographic distance to servers
  * Network congestion and routing
  * Internet service provider performance
  * CDN and edge server optimization
</Card>

<Card title="Processing Load" icon="microchip">
  **System Performance**

  * Server capacity and utilization
  * Concurrent request handling
  * Model loading and initialization
  * Resource allocation efficiency
</Card>

<Card title="Audio Quality" icon="volume-high">
  **Input Characteristics**

  * Audio sample rate and quality
  * Background noise levels
  * Speaker clarity and volume
  * Connection stability
</Card>

### Best Practices

<AccordionGroup>
  <Accordion title="ASR Optimization">
    **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
  </Accordion>

  <Accordion title="Model Selection">
    **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
  </Accordion>

  <Accordion title="TTS Configuration">
    **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
  </Accordion>
</AccordionGroup>

***

## Next Steps

<Check>
  **Ready to optimize your latency?** Use this comprehensive guide to select the optimal configuration for your specific use case and performance requirements.
</Check>

**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

<Info>
  **Need help with configuration?** Contact [support@vodex.ai](mailto:support@vodex.ai) for personalized latency optimization assistance.
</Info>

***

## Related Configuration Settings

Configure these latency-related settings in your Vodex dashboard:

<Card title="ASR Configuration" icon="microphone">
  **Speech Recognition Settings**

  Configure ASR providers and settings in [Call Settings](/call-settings/overview) and [Advanced Settings](/call-settings/advanced-settings).

  * Choose between streaming and non-streaming ASR
  * Configure language codes and detection
  * Set up custom ASR parameters
</Card>

<Card title="AI Model Selection" icon="brain">
  **Model Configuration**

  Select and configure AI models in [Call Settings](/call-settings/overview).

  * Choose from Vodex optimized models
  * Configure OpenAI, Llama, and other providers
  * Balance latency vs capability requirements
</Card>

<Card title="Voice/TTS Settings" icon="volume-high">
  **Text-to-Speech Configuration**

  Configure voice providers and settings in [Call Settings](/call-settings/overview).

  * Select ElevenLabs Turbo or Flash series
  * Configure alternative TTS providers
  * Optimize voice quality vs speed
</Card>