How Tenms works · Docs

How the pieces fit: the in-process SDK, Tenms Edge, the ElevenLabs bridge and the control plane, and the measured latency of each stage.

Your voice agent (Pipecat, LiveKit, custom)
  ├─ in-process SDK (Python wheel, C library) ── call summary ──► control plane
  └─ WebSocket ──► Tenms Edge  /v1/stream ──────── call summary ──► control plane

Browser or device (ElevenLabs client)
  └─ WebSocket + session token ──► Tenms Edge  /v1/elevenlabs/stream ──► ElevenLabs Agents

Your backend (holds the license)
  └─ POST /v1/session-tokens ──► control plane (https://app.tenms.cloud)

control plane ──► dashboard: calls, failure causes, interruptions, regions

System

Crate	Role	Depends on
`tenms-proto`	Wire types: `Module`, `ModuleSet`, `Claims` + `sign`/`verify`, `CallSummary`, `CallBatch`, Edge `ClientMessage`/`ServerEvent`	—
`tenms-core`	Engine, DSP, model runtime, vendor licensing, C ABI	proto
`tenms-edge`	Streaming service, ElevenLabs bridge, telemetry, revocation	core, proto
`tenms-control`	Licenses, session tokens, ingest, reads, stats, dashboard	proto
`tenms-models`	Open-model assist layer: `AssistedEngine` = `Engine` + Silero VAD / Smart Turn / CAM++ via ONNX Runtime, fed through the engine’s assist API	core, proto, `ort`
`tenms-cli`	Offline WAV runner and benchmark (`process --models`)	core, proto, models (feature, default on)
`tenms-py`	Python extension `tenms._native`	core, proto

Audio never waits on the network inside Tenms: licenses are verified offline (Ed25519), summaries are uploaded asynchronously, and the revocation list is polled in the background.

Latency budget

Stage	Budget / target	Measured	Source
Frame size	10 ms	—	SPEC
Algorithmic delay (Focus)	≤ 10 ms total, target 5 ms at 16 kHz	5 ms at 16 kHz; 7 ms at 24/48 kHz; 8 ms at 8 kHz; 0 without Focus	core README (`tests/focus.rs`)
Chunk buffering (SDK / Edge)	—	0 for whole 10 ms chunks, else < 10 ms once	python README, elevenlabs.md
Engine compute per frame, p99	< 1 ms (product target < 150 µs)	46 µs DSP with reference, 95 µs DSP + GRU H=128 (16 kHz, offline)	core README
		100 µs p99 (p50 21 µs) live Pipecat-over-Daily	`python/tests/live_reports/pipecat_daily_latest.json`
		~50-65 µs p99 on production Edge	deployment smoke test
ElevenLabs bridge processing	—	p99 ≤ 150 µs per chunk	elevenlabs.md
Network to Edge	deploy near callers	~118 ms RTT India → Singapore	deployment smoke test
`turn_prediction` after caller stops	—	~150-300 ms	python README
`turn_end` after caller stops	—	live p50 360 ms, p95 610 ms	live report
Barge-in decision, no transcripts	—	~1.0-1.3 s from onset; one continuous burst needs 800 ms of voicing	core README (`tests/verify.rs`), deployment facts
Barge-in decision, interim transcripts	—	~355 ms from onset	core README
Bridge gate decision after onset	—	other speaker ≈ +90 ms, noise ≈ +160 ms, barge-in ≈ +260 ms	elevenlabs.md