nVoq Voice

Dictation topics are documented as U.S. English, and public evidence does not establish ambient performance across accents, languages, overlapping speakers, speech disabilities, cognitive impairment, low-volume speech, noisy homes, or interpreter-mediated visits. The Windows guide recommends no more than four speakers and avoiding overlap. Vendor claims bias testing and fairness monitoring, but no subgroup results or patient-partnered evaluation are public.

nVoq says customer data is tenant-isolated, encrypted, regionally domiciled, covered by SOC 2 Type II controls, and not used with PHI to train its models. It says third-party generative-AI vendors such as OpenAI use zero-data-retention APIs and cannot train on customer data. Current Voice Assistant documentation also says audio, transcripts, and summaries are stored locally during processing and remain in nVoq cloud storage, with group-configured retention. This limits any blanket zero-retention interpretation. Patient access, export, deletion requests, feedback-data handling, and exact retention periods are not publicly established.

nVoq repeatedly assigns accuracy responsibility to the clinician and requires review before EHR posting. The system can nevertheless shape consequential hospice narratives, completeness checks, and planned OASIS/HOPE fields that affect care coordination, eligibility, audits, and reimbursement. Public materials do not document patient-safety escalation, hallucination rates, or independent ambient-summary accuracy.

Published evidence centers on nVoq and Amedisys quality-improvement work measuring narrative length, required hospice elements, error rates, documentation time, clinician satisfaction, and reimbursement readiness. The studies are useful operational evidence but are vendor/customer produced; no independent patient-outcome, patient-experience, agency, subgroup-fairness, or ambient-summary validation was identified.