# Latency Optimization

> Reduce response times

Source: https://edesy.in/docs/voice-agent/troubleshooting/latency

Minimize delay between user speech and agent response.

## Understanding Latency

Total latency = STT + LLM + TTS + Network

| Component | Typical Latency |
|-----------|-----------------|
| STT | 100-500ms |
| LLM | 200-1000ms |
| TTS | 100-500ms |
| Network | 50-200ms |
| **Total** | **450-2200ms** |

## Optimization Strategies

### 1. Use Native Audio Models

Bypass STT/TTS entirely:

```json
{
  "llmProvider": "gemini-live-2.5"
}
```

| Model | Avg Latency |
|-------|-------------|
| Traditional Pipeline | ~1200ms |
| Gemini Live 2.5 | ~400ms |
| OpenAI Realtime | ~500ms |

### 2. Use Vertex AI (Gemini Live)

Vertex AI reduces Gemini Live latency significantly:

```json
{
  "llmConfig": {
    "gemini-live-2.5": {
      "vertexai": {
        "enabled": true,
        "region": "us-central1"
      }
    }
  }
}
```

### 3. Choose Fast Providers

**Fastest STT:**
- Deepgram Nova-2

**Fastest TTS:**
- Cartesia
- Deepgram
- Sarvam

**Fastest LLM:**
- Gemini 2.5 Flash-Lite

### 4. Optimize LLM Response

```json
{
  "llmConfig": {
    "maxTokens": 150,
    "temperature": 0.5
  }
}
```

Shorter responses = faster TTS.

### 5. Reduce Network Latency

- Use nearest region
- Check for network issues
- Consider dedicated connection

## Measuring Latency

Check latency in Call History:

1. Open a call
2. View **Latency Breakdown**
3. See per-component timing

## Latency Targets

| Use Case | Target |
|----------|--------|
| Sales calls | <800ms |
| Support | <1000ms |
| IVR | <500ms |

## Quick Wins

1. Switch to Gemini Live
2. Enable Vertex AI
3. Use Deepgram STT
4. Keep responses short