> ## 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.

# How It Works

> SLNG runs three stages between your orchestrator and the models, one for each part of the voice pipeline. Each removes a category of unnecessary compute.

SLNG runs three stages between your orchestrator and the models you use, one for each part of the voice pipeline. Each removes a category of unnecessary compute.

## Request lifecycle of a single turn

When a caller speaks:

1. **Audio arrives** from the caller via your orchestrator.
2. **STT Routing** selects the transcription model based on language, accent, noise profile, and cost constraints.
3. **Transcript** goes to your orchestrator's LLM, or to SLNG's tiered decisioning.
4. **Tiered Decisioning** determines whether the turn needs full inference, local inference, or can be resolved without calling the LLM.
5. **The response** (LLM, cached, or deterministic) is sent to TTS.
6. **Output Assembly** checks cache, assembles from segments where possible, and generates only what is new.
7. **Audio returns** to the caller.

These stages do not have to run in strict sequence. Assembly can begin before the LLM has finished its full response.

## Three stages

### STT Routing `PRIVATE BETA`

The first stage. When audio arrives, it is routed to the right STT model for that specific interaction. A Hindi caller in Mumbai gets a different model than an English caller in New York. A noisy environment may route to a model with better noise handling. The layer balances accuracy, latency, and cost per turn.

See [STT Routing](/execution-layer/stt-routing) for details.

### Tiered Decisioning

The second stage. Not every turn needs full LLM inference. A consent disclosure repeated for the hundredth time does not need a reasoning model. It needs a deterministic, low-latency response.

Intelligence is allocated where it is needed and skipped where it is not.

See [Tiered Decisioning](/execution-layer/tiered-decisioning) for details.

### Output Assembly

The third stage. TTS on the execution layer is assembly, not generation. When a turn produces text for speech, the layer checks whether this output, or parts of it, has been assembled before. A miss is synthesized, cached, and served from cache next time.

See [Output Assembly](/execution-layer/output-assembly) for details.

## Regional execution

All three stages run across a global edge network. By default, requests route to the lowest-latency region. You can pin to specific regions for compliance or performance.

See [Regional Execution](/region-override) for details.

## Continuous improvement

As more calls flow through the layer, routing decisions sharpen, cache coverage grows, and tier allocation becomes more accurate. Every call makes the next one more efficient.