FormBharo：为印度农村地区对话式填表而设计与评估的语音代理

文章背景与核心概要

在印度农村等许多发展中地区，获取关键的社会福利和医疗项目往往从填写表格开始。然而，目标受益人经常面临读写能力不足的问题，这使得语音对话式交互成为刚需。传统上，这一重任落在基层卫生工作者的肩上，他们需要逐一为个人进行注册——考虑到他们有限的工作能力，这构成了主要的业务瓶颈。

为了应对这一挑战，研究人员开发了 FormBharo（在印地语中意为“填表”），这是一种创新的语音代理，旨在严格的延迟和成本限制下通过电话完成结构化表格的填写。该代理将大语言模型（LLMs）与确定性的、基于规则的验证及流程控制相结合。目前，FormBharo 正与非政府组织 ARMMAN 合作进行试点，用于妇幼移动医疗项目，协助低收入、讲印地语的母亲注册产前和产后护理服务。

📌 摘要

In many developing regions like rural India, accessing vital social benefits and healthcare programs begins with filling out forms. However, target beneficiaries are frequently unable to read or write, necessitating a spoken conversational approach. Traditionally, this responsibility falls onto frontline health workers who enroll individuals sequentially—a major bottleneck given their stretched capacities.

在印度农村等许多发展中地区，获取关键的社会福利和医疗项目往往从填写表格开始。然而，目标受益人经常面临读写能力不足的问题，这使得语音对话式交互成为刚需。传统上，这一重任落在基层卫生工作者的肩上，他们需要逐一为个人进行注册——考虑到他们有限的工作能力，这构成了主要的业务瓶颈。

To address this challenge, researchers developed FormBharo (meaning "fill the form" in Hindi), an innovative voice agent designed to complete structured forms over phone calls under strict latency and cost constraints. The agent pairs Large Language Models (LLMs) with deterministic, rule-based validation and flow control. Currently piloted alongside the NGO ARMMAN for maternal and child mobile-health programs, FormBharo assists low-income, Hindi-speaking mothers in enrolling for antenatal and postnatal care.

为了应对这一挑战，研究人员开发了 FormBharo（在印地语中意为“填表”），这是一种创新的语音代理，旨在严格的延迟和成本限制下通过电话完成结构化表格的填写。该代理将大语言模型（LLMs）与确定性的、基于规则的验证及流程控制相结合。目前，FormBharo 正与非政府组织 ARMMAN 合作进行试点，用于妇幼移动医疗项目，协助低收入、讲印地语的母亲注册产前和产后护理服务。

🔍 核心发现与贡献

First-of-its-Kind Pilot: To the authors' knowledge, FormBharo is the first voice agent deployed to handle enrollment form-filling for this specific population demographic.

同类首创试点： 据作者所知，FormBharo 是第一个针对该特定人口群体部署的、用于处理注册填表任务的语音代理。

FormVoiceAgentBench Benchmark: The team openly released a comprehensive benchmark pairing human-recorded Hindi audio with 3,760 multi-turn conversation tests across 960 simulated calls. This framework evaluates component performance (transcription, extraction, reply generation) and end-to-end form completion under real-world acoustic variations.

FormVoiceAgentBench 基准测试： 该团队公开了一个全面的基准测试，将人工录制的印地语音频与 960 次模拟通话中的 3,760 项多轮对话测试进行了配对。该框架评估了在真实世界声学变化下的组件性能（转录、提取、回复生成）以及端到端表格完成情况。

Impact of Real-World Speech: Form completion accuracy drops by up to ~41 points when LLMs are fed error-prone real-speech transcripts instead of clean, reference transcripts.

真实世界语音的影响： 当向大语言模型输入容易出错的真实语音转录文本，而非干净的参考转录文本时，表格完成准确率最多下降了约 41 个百分点。

The Value of Rule-Based Controls: Deterministic, rule-based controls successfully recover many turn-level extraction errors, enabling smaller and cheaper models to perform on par with—or even surpass—frontier models in overall form completion.

基于规则控制的价值： 确定性的、基于规则的控制成功纠正了许多轮次级别的提取错误，使得更小、更便宜的模型在整体表格完成度上能够与前沿模型相媲美，甚至超越它们。

Pipeline Discrepancies: Component performance does not accurately predict end-to-end success. For instance, while GPT-5.5 leads in turn-level extraction accuracy on reference transcripts (99.8%), it ranks lower on actual end-to-end form completion because errors propagate and cancel out across the pipeline.

流水线差异： 组件性能并不能准确预测端到端的成功。例如，尽管 GPT-5.5 在参考转录文本的轮次级别提取准确率上处于领先地位（99.8%），但在实际的端到端表格完成度上排名却较低，因为错误会在整个流水线中传播并互相抵消。

Deployment Optimization: Since no single model excels simultaneously in accuracy, cost, and latency, the authors utilize a Pareto-based weighted-sum scalarization technique to determine the optimal deployable model configuration.

部署优化： 由于没有任何单一模型能在准确性、成本和延迟方面同时表现卓越，作者采用了一种基于帕累托（Pareto）的加权和标量化技术来确定最佳的可部署模型配置。