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

# What it optimizes

## Listen

**Cleaner input audio**

Noise cancellation runs on the inbound stream before transcription. Cleaner audio means lower Word Error Rate, which means fewer turns where the model is reasoning over the wrong words.

**Model choice or your own key**

Use any model in the SLNG catalogue, or bring your own provider key, without changing how your agent calls the Listen step.

## Think

**Per-turn routing**

Each request is evaluated before it goes anywhere. Turns that can be served from cache, from a prepared response, or from a small scoped model are, and only the turns that need frontier reasoning get it.

**Lower latency**

Turns served from cache or from a local in-region model return in a fraction of the time a round trip to a frontier provider takes.

**Predictable output**

Consistency comes from routing, not from prompting. Every turn answered without open-ended frontier composition is a turn without the variance that composition introduces.

**Centralised model configuration**

Model selection, your own provider keys, and per-call personalisation live in one configurable place instead of being scattered across your integration code. Changing provider is a configuration change.

**Automatic fallback**

If a configured provider fails or degrades, the request is routed to an alternative without your agent handling the retry.

**Observation mode**

Point the router at your real traffic without repointing your agent. It records what would have happened and reports the routing and coverage it would have achieved, so you can validate behaviour before you switch over.

See [Context Router](https://slng-new-docs.mintlify.site/guides/execution-layer/llm/context-router) for configuraion and routing behaviour in detail.

## Speak

**Lower time to first audio**

Cached audio is served from our region based platform and not synthesized on demand. This removes the synthesis wait at the point in the turn where silence is most audible to the caller.

**Repeatable output**

Phrases your agent says on every call are served as the same audio every time, so recurring lines don't vary between calls.

**Region, Scoping and retention you control**

PII aware cache at project and organisation level. Anything you're re using is saved in Region and focused on compliance.

**Pronunciation dictionaries**

Define how names, product terms, and domain vocabulary are spoken, so pronunciation is consistent across every voice and every call.

## Across the layer

**One integration surface**

Listen, Think, and Speak are configured through the same layer. Adding, removing, or changing a capability at any step does not change how your orchestrator calls it.

**Regional execution**

The layer runs in-region across the SLNG cluster footprint, so optimisation happens close to the caller rather than adding a hop to a distant control plane.

## The system gets better the more you use it

Every call through the Execution Layer sharpens routing decisions and widens cache coverage for the next one. Latency drops. Control and reliability go up. Simple as that.

* More calls, more cache coverage, fewer model calls, more control
* More patterns observed, better routing decisions, lower latency
* More providers configured at the Think step, more failover options, higher reliability