Voice-First Health Guidance in Hindi and Bengali: a Retrieval-Grounded Speech-To-Speech Assistant with a Verified-Response Cache and a Preliminary Evaluation

India has about half the density of doctors, nurses and midwives that the World Health Organization recommends, and most digital health tools assume a user who reads English. We describe a prototype speech-to-speech assistant that gives preliminary health guidance in Hindi and Bengali. Spoken input is transcribed by a commercial speech recognition service, classified by a large language model, and then routed to one of two paths: a cache of pre-verified answers for frequent intents, or retrieval-augmented generation over a medical knowledge base. We report a preliminary evaluation and state its limits. On 10 synthetic queries per language scored by a language model acting as judge, mean medical safety was 4.0 of 5 in Hindi (95% confidence interval 3.42 to 4.58) and 4.1 in Bengali (3.47 to 4.73); empathy and clarity were 0.3 points lower in Bengali than in Hindi, differences that were not statistically significant in a sample of this size. Four medical professionals rated diagnostic safety at 4.25 of 5, with a confidence interval of 3.57 to 4.93 that reflects the small panel. The cached path answered in 0.05 s against 1.6 s for retrieval and generation, a 97% reduction for that component only; we show that the saving across all queries is proportional to the cache hit rate, which was not measured. Word error rate rose from 8.2% to 18.1% in Hindi and from 9.1% to 19.4% in Bengali under simulated background speech. The query generator, the answering model and the judge belong to one model family, so the automated scores are indicative only. The results support a controlled evaluation with independent judges and clinicians; they do not establish clinical safety.