Appendix B.3 — AI voice

Reach respondents who don’t have internet or smartphones.

The agent calls respondents, asks follow-ups, clarifies ambiguous answers, and supports 40+ languages — including Arabic, Swahili, Hausa, Amharic, and Somali.

• Standard PSTN inbound and outbound — respondents use any phone
• Azure Communication Services + Pipecat orchestration
• GPT-5 Realtime Voice — natural turn-taking, not IVR
• Whisper-3 transcription in 40+ languages, mid-call language switch
• Claude Opus 4.7 fallback reasoning for ambiguous answers
• ElevenLabs voices for warm, localised delivery
• Consent recorded at call start, transcript retention policy per workspace
• Responses flow into the same classification + reporting pipeline as web/PWA