Bhavik Vachhani

Bhavik Vachhani, Kush Shrisvastava, Pranshu Nema et al.

Evaluating large language models (LLMs) for clinical documentation tasks such as SOAP note generation remains challenging. Unlike standard summarization, these tasks require clinical abstraction, normalization of colloquial language, and medically grounded inference. However, prevailing evaluation methods including automated metrics and LLM as judge frameworks rely on lexical faithfulness, often labeling any information not explicitly present in the transcript as hallucination. We show that such approaches systematically misclassify clinically valid outputs as errors, inflating hallucination rates and distorting model assessment. Our analysis reveals that many flagged hallucinations correspond to legitimate clinical transformations, including synonym mapping, abstraction of examination findings, diagnostic inference, and guideline consistent care planning. By aligning evaluation criteria with clinical reasoning through calibrated prompting and retrieval grounded in medical ontologies we observe a significant shift in outcomes. Under a lexical evaluation regime, the mean hallucination rate is 35%, heavily penalizing valid reasoning. With inference aware evaluation, this drops to 9%, with remaining cases reflecting genuine safety concerns. These findings suggest that current evaluation practices over penalize valid clinical reasoning and may measure artifacts of evaluation design rather than true errors, underscoring the need for clinically informed evaluation in high context domains like medicine.

Bhavik Vachhani, Chitralekha Bhat, Sunil Kopparapu

Phone level localization of mis-articulation is a key requirement for an automatic articulation error assessment system. A robust phone segmentation technique is essential to aid in real-time assessment of phone level mis-articulations of speech, wherein the audio is recorded on mobile phones or tablets. This is a non-standard recording set-up with little control over the quality of recording. We propose a novel post processing technique to aid Spectral Transition Measure(STM)-based phone segmentation under noisy conditions such as environment noise and clipping, commonly present during a mobile phone recording. A comparison of the performance of our approach and phone segmentation using traditional MFCC and PLPCC speech features for Gaussian noise and clipping is shown. The proposed approach was validated on TIMIT and Hindi speech corpus and was used to compute phone boundaries for a set of speech, recorded simultaneously on three devices - a laptop, a stationarily placed tablet and a handheld mobile phone, to simulate different audio qualities in a real-time non-standard recording environment. F-ratio was the metric used to compute the accuracy in phone boundary marking. Experimental results show an improvement of 7% for TIMIT and 10% for Hindi data over the baseline approach. Similar results were seen for the set of three of recordings collected in-house.