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

# Latency & Optimization

> How Movoice AI achieves sub-500ms response times in the Indian market.

# Latency & Optimization

In Voice AI, every millisecond counts. If an agent takes too long to respond, the conversation feels robotic and turn-taking breaks down. Movoice AI is engineered for **ultra-low latency (300ms–500ms)** — one of the fastest platforms for Indian deployments.

## How We Achieve Sub-500ms Latency

### 1. Regional Edge Servers

Most Voice AI platforms route traffic through global hubs in the US or Europe. Movoice AI routes calls via localized server clusters in **Mumbai** and **Chennai**.

* Reduced packet travel distance for Indian users shaves off 200–400ms of physics-level latency.
* Indian callers get responses processed in-region, not via Singapore or US-East servers.

### 2. Full Streaming Pipeline

We don't wait for a complete sentence before responding. Every stage of the pipeline streams in parallel:

| Stage                    | How we stream                                           |
| :----------------------- | :------------------------------------------------------ |
| **ASR (Speech-to-Text)** | Begin processing while the caller is still speaking     |
| **LLM (Reasoning)**      | Start generating text as the last word is transcribed   |
| **TTS (Text-to-Speech)** | Play the first words while the rest is still generating |

This triple-streaming approach compounds — cutting 300–500ms versus a sequential pipeline.

### 3. Predictive Turn-Taking

Movoice AI uses **Voice Activity Detection (VAD)** to distinguish between a natural pause (breathing) and the end of a user's thought, minimizing dead air between turns.

***

## Optimizing Your Agent

You can further reduce latency with these configuration choices:

| Setting              | Recommendation                            | Impact                               |
| :------------------- | :---------------------------------------- | :----------------------------------- |
| **LLM Model**        | Use Gemini 1.5 Flash or GPT-4o mini       | Fastest inference for routine tasks  |
| **Voice**            | Select voices tagged `Low Latency`        | Sarvam AI voices are optimized       |
| **Webhook Response** | Keep your tool call responses under 100ms | Slow tools are the #1 latency killer |
| **Prompt Length**    | Keep system prompt under 800 tokens       | Shorter context = faster first token |

***

## Latency Benchmarks

| Platform                | Response Time (Global) | Response Time (India) |
| :---------------------- | :--------------------- | :-------------------- |
| **Legacy IVR**          | 2s – 5s                | 2s – 5s               |
| **Global AI Platforms** | 600ms – 900ms          | 1.2s – 1.5s           |
| **Movoice AI**          | **400ms – 600ms**      | **300ms – 500ms**     |

<Note>
  To maintain these speeds, use a stable internet connection for web calls or high-quality SIP trunks for telephony. Unstable networks can add 100–200ms regardless of server proximity.
</Note>

***

## Troubleshooting Slow Responses

**Agent feels slow?** Check these in order:

1. **Model** — switch to Gemini Flash if using GPT-4 Turbo
2. **Voice** — avoid voices not marked `Low Latency`
3. **Tool calls** — if your agent hits an external API, that response time adds directly to latency
4. **Prompt** — a 3,000-token system prompt is significantly slower than a 600-token one
5. **Network** — test from a different network to rule out local connectivity issues