Intelligent IVR Voice Interaction: Architecture, Models, and Deployment at Ant Financial

Interactive Voice Response (IVR) is a telephone‑based system that lets customers navigate menus using keypad presses or voice input. Traditional IVR suffers from complex menus, long wait times, and hidden human‑service entry points, which limit user satisfaction and increase operational costs.

To address these issues, Ant Financial built an intelligent IVR platform named MISA (Machine Intelligence Service Assistant). By integrating speech recognition, machine learning, and natural language understanding, MISA enables users to speak their requests, provides personalized guidance, and routes calls to the most suitable service agents, dramatically improving experience and reducing costs.

MISA’s core functions include: (1) guessing the user’s problem and pushing solutions via the Alipay app; (2) routing unresolved issues to the appropriate skill group; and (3) transferring user‑provided context to human agents to accelerate resolution.

The interaction engine behind MISA uses deep learning and NLU techniques to deliver personalized, intelligent, fine‑grained, and configurable dialogues. It supports dynamic configuration and A/B testing of prompts.

Key technical components:

Guess‑question: a hybrid network combining CNN, LSTM, and ANN processes service‑trajectory, textual, and factor features to predict user intent before the user speaks, achieving high coverage and accuracy.

Problem identification: a two‑stage model (sub‑model + fusion) that leverages semantic matching, intent‑tree analysis, search‑engine fallback, and a hierarchical element model, with a GBDT layer to fuse results.

Reverse questioning: a multi‑turn dialogue strategy that decomposes queries into three elements (business, framework, problem type) to generate concise follow‑up questions, keeping interactions within three turns.

Automated training and deployment: daily model updates using internal platforms (核能, armor) for feature processing, GPU training, A/B testing, and seamless rollout, ensuring model freshness amid rapidly changing business scenarios.

During the 2023 Double‑11 shopping festival, MISA improved account‑recognition coverage by 16.6 pp and accuracy by 6.4 pp, problem‑identification coverage by 30.1 pp and accuracy by 15.2 pp, and the new guess‑question capability covered 42.7 % of traffic with a 63.5 % click‑through rate and 84 % accuracy.