Workflow Automation · Listen + Speak
IVR → STT: smart voice routing
Automate
Rep · High
Reason · Low
The problem
Old IVR menus force callers through "Press 1 for billing…". Drop-off is high; agents still receive mis-routed calls.
The AI approach
Open-ended STT + intent classifier: caller says "I want to update my address" — speech is transcribed, intent recognised, call routed.
The outcome
−40s avg.
Setup falls ~80s → ~40s. Mis-route rate 12% → 2%. ~30% self-serve.
Try itInput → Process → Output
Input — Inbound call
Caller · 089-***-4421
picked up · 0:02
"I want to update my address"
Process — AI pipeline
1STT — speech to textListenPredictive
2Classify intent + entityReadPredictive
3Confidence + route rulesRulesSymbolic
4TTS reply (self-serve)SpeakGenerative
Output — Routing decision + reply
Click Run demo to see speech turned into a routing decision and a spoken reply.
Transcript · 0.94 conf.
"I want to update my address"
Intentupdate_address0.94
Route→ Self-serve≥ 0.85
TTS · spoken back
"Sure — what's the new address?"
Three AI types in this use case
SymbolicConfidence threshold, routing tree (intent → queue), instant-bail keywords, PII redaction rules.
PredictiveSTT acoustic model + intent classifier (top-12 intents trained on labelled history).
GenerativeLLM disambiguates ambiguous utterances; TTS synthesises the self-serve reply.
The stack
- STT · Whisper-large / Deepgram Nova
- Intent · LLM classifier (few-shot)
- Telephony · Twilio / Genesys
- TTS · ElevenLabs / Azure Neural
When this works
- Top intents are ≤ 12 and well-defined
- Audio over a clean carrier (PSTN-quality)
- Self-serve genuinely possible (lookups, status)
- Volume justifies per-min STT cost
When it fails
- Heavy local dialect — STT WER spikes
- Low-confidence intent never escalates → loops
- PII spoken in the open — must redact before logging
- "I want a manager" reasoned away — keep instant-bail keyword