Zonov Lipi: Speech Recognition for Code-Switched Clinical Conversations in English, Hindi and Hinglish

Zonov Lipi is a 1.55-billion-parameter speech-to-text (STT) model developed by Zonov.ai Labs for clinical conversations involving English, Hindi, and Hinglish. It is designed to address the challenges of Indian medical terminology, code-switched speech, and accurate clinical transcription. This research presents a safety-oriented evaluation framework that measures word error rate (WER), critical errors involving drug names and dosages, drug-name recall, and hallucinations on non-speech audio. The model is evaluated against commercial and open-source speech recognition systems across ten test sets containing 2,155 audio clips. The reported results demonstrate improvements in Hinglish transcription and drug-name recognition compared with the evaluated baseline systems. The report also discusses limitations, including hallucinations on noise and music, lower accuracy on some monolingual speech benchmarks, and the need for clinician review of medical transcripts. Zonov Lipi is designed to support on-premises deployment, allowing healthcare organizations to process audio within their own infrastructure. This technical report describes the model's architecture, training methodology, benchmarking results, security controls, privacy considerations, and future development plans. Organization: Zonov.ai LabsResearch Areas: Healthcare AI, Automatic Speech Recognition, Medical NLP, and AI Governance.