Glossary - SLNG: the execution layer for voice - SLNG Documentation

Short definitions for the terms that appear across these docs. Each term links to the page that covers it in full.

Products

Term	Definition
Execution Layer	The optimisation layer between your orchestrator and the models. It applies caching, routing, and quality features to the Listen, Think, and Speak steps without changing how your agent calls them.
Model Gateway	One low-latency entry point that routes a speech-to-text or text-to-speech request to the right model in the region you call.
Context Router	The component in front of the Think step. It decides per turn whether to answer from cache, a small in-region model, a prepared response, or the full model.
Unified API	A normalisation layer so every speech provider uses the same request and response shape. Switching providers is a change to the model name, with no other code changes.
Agent Builder	Where you create, configure, and run voice agents from the dashboard and API.
Unmute	An open declarative standard for voice agents: describe an agent once, then compile it to a target such as LiveKit, Pipecat, or SLNG.

Building blocks

Term	Definition
Project	The organizing unit. Everything one voice use case needs (agent, models, keys, tools, telephony, calls) lives inside its project.
Agent	A voice agent: a configuration that pairs speech-to-text, language, and text-to-speech models with instructions and tools to hold a conversation.
Listen, Think, Speak	The three steps of a voice agent, mapping to speech to text, the language model, and text to speech. Each is configured and optimised on its own.
Orchestrator	The software that runs the conversation and glues the models together, either LiveKit or Pipecat. It owns turn-taking and calls the Execution Layer rather than the providers directly.
Pipeline	The speech-to-text → language model → text-to-speech chain a request passes through on every turn.

Capabilities

Term	Definition
Speech to text (STT)	The Listen step. Turns spoken audio into written text as it happens.
Text to speech (TTS)	The Speak step. Turns written text into spoken audio.
Large language model (LLM)	The Think step. Reads the transcript and decides how the agent responds.
ASR	Automatic speech recognition. Another name for speech to text, used by some providers.

Regions and infrastructure

Term	Definition
Region	One of the 14 geographic locations a workload runs in. Each region has its own endpoint, and requests are not routed between regions.
In-region execution	Running the whole request path close to the caller, in the same region, so no hop is added to a distant location.
Data residency	Keeping voice data inside a chosen region to meet the rules of a jurisdiction or compliance framework.
PII	Personally identifiable information. Voice is treated as PII under several compliance frameworks, which is why region and retention are controllable.

Models and access

Term	Definition
Provider	The vendor of a model, such as Deepgram or Cartesia. A provider can be SLNG-hosted or reached with your own key.
SLNG-hosted vs proxied	An SLNG-hosted model runs in-region on SLNG infrastructure. A proxied model is an external provider reached through the gateway.
BYOK	Bring your own key. Use your own provider keys through SLNG instead of SLNG-hosted models.
Model ID	The identifier that selects a model, in the form `provider/family:version-lang`. The `slng/` prefix selects an SLNG-hosted, in-region route; a bare `provider/model` is the external or BYOK route.

Conversation quality and latency

Term	Definition
Turn	One exchange in a conversation: the caller speaks, the agent responds. Turn-taking is real time and has a hard latency budget.
Latency	The delay before the agent responds. Distance to the region is the largest lever, ahead of any model or pipeline change.
Time to first audio (TTFB)	How long the caller waits before hearing any audio on a turn. Cached audio removes the synthesis wait.
Word error rate (WER)	How often the transcription gets a word wrong. Cleaner input audio lowers it, so the model reasons over the right words.
VAD	Voice activity detection. Detects when the caller is speaking, which drives turn-taking.
Endpointing	Deciding when a caller has finished a turn, so the agent can respond without cutting them off.
Diarization	Separating a transcript by speaker, so it is clear who said what.
Partials vs finals	In streaming transcription, a partial is an interim result that can still change; a final is the settled text for that segment.

Execution Layer optimizations

Term	Definition
Noise cancellation	Cleans the inbound audio before transcription, which lowers the word error rate.
Pronunciation dictionary	A set of rules for how names, product terms, and domain words are spoken, applied across every text-to-speech provider.
TTS caching	Reuses previously synthesized audio for phrases an agent repeats, cutting cost and time to first audio.
Fallback	An ordered list of backup models. If a provider fails or times out, the request moves to the next one without your agent handling the retry.
Observation mode	Points the Context Router at your real traffic without repointing your agent, recording the routing and coverage it would have achieved.
Voice cloning	Building a custom text-to-speech voice from audio samples, on providers that support it.