LiveKit NCRI Transcription Benchmarks

Summary

10 concurrent NCRI silent-transfer calls on shared VM 194.146.13.204 (12 cores, ~47 GB). Path: ESL → DN 10005 → LiveKit NCRI-agent → FusionPBX/EFCX (99900) with dual-leg STT → CCM.

Transfers worked (10/10). Dual STT did not scale to 10 because ElevenLabs caps 15 concurrent STT streams and NCRI uses 2 streams per call (safe ≈ 7 answered calls). First infra CPU ceiling was livekit-sip.

Test window: 2026-08-11 ~10:37–10:40 UTC. CIDs: 3586444663–3586444672.

Results

Metric	Value
Setup	10/10 originates OK
Active at peak	10/10 for ~75s
First >80% CPU	`livekit-sip` @ 111% with 10 active calls
Peak `livekit-sip`	~255% (Docker multi-core %)
Peak `livekit-server`	~52%
Redis / FreeSWITCH / NCRI	Not the ceiling

ElevenLabs STT limit

Fact	Value
Max concurrent STT	15
Streams per NCRI call	2 (customer + agent)
Safe concurrent answered calls	~7
At 10 calls	Up to 20 streams → `rate_limited`

Example: CID 3586444665 answered and CALL_CONNECTED OK; transcription failed due to STT rate limit (not transfer/SIP/desk).

Next steps

Raise ElevenLabs concurrent STT quota (required for >7 dual-STT calls).
Scale out livekit-sip off the shared 12-core box.
Re-benchmark after both before targeting higher concurrent NCRI load.

Repo reference: agent/docs/NCRI_10_CALL_BENCHMARK.md (branch main / develop, tip da672d1).