What Is Contact Center PSTN (Public Switched Telephone Network)?

What Is Contact Center PSTN (Public Switched Telephone Network)?

Contact Center PSTN (Public Switched Telephone Network) is the public phone-network edge that contact centers use to receive and place calls across landline and mobile carriers. It includes legacy circuit-switched routes and modern SIP/IP interconnects, so codec, packet-loss, jitter, carrier, and country differences can change what an LLM voice agent hears. FutureAGI treats that context as part of evaluation because PSTN artifacts explain why the same LiveKit or Pipecat agent passes internal WebRTC tests but regresses on real customer calls.

Why Contact Center PSTN matters in production LLM and agent systems

Many “the bot got worse on calls from carrier X” reports are PSTN-side problems. Different carriers negotiate different codecs at the SIP-trunk edge. Mobile-originated calls to the same number have different jitter profiles than landline calls. International calls add 80–150 ms of one-way latency and break turn-taking models tuned for North American RTT. Calls routed through a different SBC during failover lose codec consistency, and ASRAccuracy regresses on a slice of traffic the team didn’t know existed.

The pain spreads across roles. A voice engineer chases a complaint that “the bot doesn’t understand customers from Australia” and finds the international SIP trunk is using G.729. A compliance officer is asked whether outbound calls to specific countries respect local consent laws and the team has no per-country routing in the gateway. An ops lead sees a 4% resolution drop on Tuesdays and traces it to a carrier that re-routes traffic through a different POP at peak hours.

In 2026 the PSTN is heterogeneous and changing fast. Carriers are migrating from TDM to SIP, codecs vary by region, and PSTN-replacement services (IMS, RCS) are emerging. Unlike a MOS-only dashboard or WebRTC-only QA run, PSTN-aware evaluation has to connect audio quality to the actual carrier route. The only way to keep an LLM voice agent stable across this mess is to expose PSTN context — codec, originating carrier, country, SIP trunk id — on every voice trace and slice eval results by it.

How FutureAGI handles the PSTN-to-LLM bridge

FutureAGI’s approach is to treat PSTN context as span attributes on every voice-agent trace, so PSTN-induced regressions can be isolated from model behavior. With traceAI-livekit or traceAI-pipecat wired, voice spans carry voice.codec, voice.session.id, and any inherited SIP fields. ASRAccuracy runs on the transcript; AudioQualityEvaluator runs on the audio leg; both can be sliced by codec, SIP trunk, and originating carrier. ConversationResolution ties the call’s outcome to the PSTN cohort.

A concrete example: a global SaaS company runs an inbound support voice agent that takes calls from 30 countries. After expanding to APAC carriers, ASRAccuracy drops on calls from Australia and Singapore. The team filters traces by voice.codec and finds the APAC SIP trunk negotiates G.729 instead of G.711. They reconfigure the trunk to prefer G.711 wideband, retrain the bot’s Australian-English ASR cohort with simulate-sdk Persona cases, and ConversationResolution for APAC inbound climbs from 0.71 to 0.83. Without PSTN-aware tracing, this looked like a model regression and would have triggered a costly model swap.

How to measure contact center PSTN quality

PSTN-bridged voice agents need PSTN-aware signals:

• ASRAccuracy by voice.codec and SIP trunk: codec mismatches and trunk-specific issues surface here.
• AudioQualityEvaluator by originating country: regional carrier quality varies; track separately.
• End-to-end latency by route: international calls add round-trip time that breaks turn-detection.
• Per-trunk fail rate: percentage of calls per SIP trunk that fail eval thresholds.
• Codec-mismatch rate: count of negotiated-codec drift events that correlate with quality drops.

Minimal Python:

from fi.evals import AudioQualityEvaluator

evaluator = AudioQualityEvaluator()
result = evaluator.evaluate(
    input="PSTN-bridged audio",
    output=audio_path,
)
print(result.score, result.reason)

Common mistakes

• Treating the PSTN as uniform. Carriers, codecs, packet loss, and jitter differ by route; trace by trunk, country, and negotiated codec.
• Logging only transcript text. If traces omit voice.codec and carrier metadata, codec failures look like ASR or model regressions.
• Tuning turn detection on lab WebRTC calls. PSTN jitter and international RTT change interruption behavior, especially for barge-in and silence timeouts.
• Skipping per-country compliance routing. Outbound rules vary by jurisdiction; encode consent checks and caller-id policy before traffic reaches the gateway.
• Using one global ASR cohort. Regional accents over narrowband PSTN audio need accent-aware eval slices, route changes, or targeted retraining.