SLNG is an execution layer that sits between a voice-agent orchestrator and the speech and language models it calls. You keep your orchestrator (LiveKit, Pipecat, or a custom pipeline), your prompts, and your tools, and repoint the model endpoints at SLNG; it then routes each turn, caches, redacts personal data, and logs what happened.
The layer has three stages: a speech-to-text performance layer, a context router that decides per turn whether a request needs the LLM, a local model, or no inference at all, and a text-to-speech path with caching. Optional STT and TTS gateways front 30+ models from providers such as Deepgram, AssemblyAI, Speechmatics, Soniox, Whisper, Gladia, Cartesia, Rime, Murf, InWorld, and Sarvam, either on SLNG's contracts or with your own keys. SLNG's published figures for adaptive execution are 39% lower turn latency, 53% lower model cost, and 16% better call outcomes.
Execution can be pinned to any of 11 regional hubs through a request header so that audio and text stay in a chosen jurisdiction. The service exposes HTTP and WebSocket APIs, ships livekit-plugins-slng and pipecat-slng adapters, and bills per agent minute with no contracts or minimums.
Features
- Drop-in execution layer: change the endpoint in your agent config while the orchestrator, prompts, and tools stay the same
- Per-turn routing: a context router picks LLM, local, or no inference for each turn
- STT gateway: multi-provider, multi-language transcription with per-turn model selection
- TTS gateway: multi-provider synthesis with caching and PII-safe delivery
- Bring your own keys: use existing OpenAI, Anthropic, Deepgram, or ElevenLabs contracts through SLNG
- Regional pinning: 11 sovereign hubs selectable per request via header
- Governance: PII redaction and logging of every routed call
- Framework adapters: livekit-plugins-slng and pipecat-slng
- Compliance: ISO 27001:2022, HIPAA, and GDPR, with SOC 2 Type II in progress
