A voice agent is filling health forms by phone in rural India

The form is the product

Most talk about voice agents starts with how natural they sound. FormBharo starts somewhere less glamorous: can a phone call produce a usable enrolment form?

The system is being piloted with ARMMAN, an Indian nonprofit, for antenatal and postnatal care enrolment in rural Maharashtra. A caller speaks in Hindi. There is nothing to read, type or install.

That is useful for people who may struggle with written forms, and for frontline health workers who currently collect the same information one conversation at a time. It also makes every error consequential. A misunderstood date or phone number is not just an awkward reply. It changes the record.

The researchers therefore judge the agent on the completed form, not on whether one model response looked convincing.

Language models get a narrow job

FormBharo is deliberately hybrid. Speech recognition first turns the call into text. One language model extracts the answers and gives a short acknowledgement. A second phrases the next question. Text-to-speech reads it back.

The form itself is controlled by ordinary rules. They check whether a phone number has ten digits, whether a date is plausible, when to retry and which branch comes next. If someone says she is pregnant, for example, the system asks about gestational age. If not, it follows a different path.

This division is not especially magical. That is the point. The models handle messy language; deterministic software decides what counts as a valid field and keeps the state of the form.

In one test, a smaller model returned a number as text. Its extraction score fell, but the rule layer normalised the value and the final form stayed correct. A component leaderboard would miss that recovery.

A benchmark built around the whole call

The team created FormVoiceAgentBench from the 12-field enrolment flow used in the pilot. It contains 3,760 multi-turn tests assembled into 960 simulated calls.

Five Hindi-speaking women recorded 380 answers. Each answer was captured in an ideal setting and under variations such as background chatter, greater microphone distance or a different speaking pace. The recordings were then combined with branching call scripts for five simulated user profiles.

No real caller data was used in the benchmark. The profiles and reference transcripts were generated with a language model under constraints agreed with ARMMAN, while people recorded the audio. That makes the test repeatable, but it is not the same as an unscripted call with a real enrollee.

The researchers compared five speech-to-text systems and 11 language models. Open-ended names and replies were partly scored by calibrated model judges; closed fields such as dates and booleans used exact matching.

Speech errors get larger downstream

On clean reference transcripts, several models looked nearly perfect at extracting individual answers. Once the benchmark used transcripts made from the recorded speech, the rankings changed.

With the three speech systems carried into the full test, median form-completion accuracy fell by 7.4, 7.7 and 13.6 percentage points compared with reference transcripts. The largest model-specific fall was about 41 points. Small recognition mistakes had accumulated across the call.

The team's selected configuration did not simply pick the most accurate model at every step. It balanced completed-form accuracy against response time and cost, ending with Scribe v2 for transcription, Gemini 3.5 Flash for extraction and GPT-5.4-mini for the reply.

Those names will date quickly. The more durable result is the method: test the chain people actually use, with the noise and budgets it will actually face.

What is confirmed, found and still open

Confirmed: the researchers submitted the FormBharo preprint on 6 August 2026. It describes an ongoing rural Maharashtra pilot with ARMMAN and a Hindi benchmark using human-recorded audio, simulated profiles and no real caller data.

The research finding: component accuracy did not reliably predict a correct final form. Rule-based checks repaired some errors, while uncorrected speech mistakes grew as they moved through the call.

The authors' claim: they describe FormBharo as the first conversational voice agent piloted for enrolment with this population. That priority claim has not been independently established here.

Still open: completion rates with real callers, consent and privacy in routine use, performance across more accents and spontaneous speech, and what happens when a caller needs a person. The paper points to a code and data repository, but that link did not resolve publicly when this article was prepared.

It is an early pilot, not proof that an agent can replace a health worker. It does show a more grounded way to build one: give the model the part that needs language, and keep the form on a shorter leash.