Miravoice

Phone-first automation that preserves classical survey methodology primitives (branching logic, skip patterns, randomization, multiple question types) — this implies a non-trivial orchestration layer that maps abstract survey designs to timed telephony dialogues, not just a simple IVR tree.

Large multilingual stack (14 languages) tuned for survey use-cases — handling short answers, single-word responses, numeric scales and free-text requires language-specific ASR/NLU models and robust turn-taking/confirmation strategies (e.g., disambiguation prompts, confidence thresholds) rather than one-size-fits-all speech-to-text.

Published, conference-grade validation (AAPOR & MAPOR 2025) — they’ve run controlled experiments comparing automated phone agents to human interviewers. That signals collection of labeled outcomes (response rates, item nonresponse, mode effects, transcription error rates) and a research pipeline for statistical validation.

End-to-end cost reduction claim (70–90% cheaper than traditional call centers) suggests optimization across several technical vectors: predictive dialing algorithms, call scheduling heuristics tuned to demographics/timezones, automated consent capture, and likely selective human-in-the-loop review to balance cost vs. quality.

Survey-design as a product primitive (randomization, ordering, skip logic) integrated with telephony — this requires an experiment engine that can perform on-call random assignment, persist respondent state across callbacks, and export structured data with provenance (timestamps, ASR confidence, audio segments).