Text-based virtual patients rehearse the content of a consultation; voice rehearses its physics, and the difference is larger than the interface change suggests. Speaking to an AI patient in real time restores the dimensions text quietly deleted, timing, interruption, hesitation, filler, the overlap of thinking and talking, which are precisely the dimensions OSCEs and real consultations run on. It also imports a set of problems text never had, accent recognition, speech differences, background noise, disability access, and an honest guide holds both columns, because the right answer is not voice versus text but a deliberate division of labour between them.
What voice genuinely adds
Five real dimensions. Timing under pressure: the consultation clock runs at speech speed, and pacing a history into eight minutes is a spoken skill that typing cannot rehearse. Interruption handling: real patients talk over, digress and return; voice simulation, where platforms implement it, trains the recovering-the-thread skill text formats structurally cannot. Hesitation and filler exposure: hearing your own "um" density and dead air is formative in a way transcripts are not, and voice practice surfaces it nightly rather than at the recorded mock. Follow-up fluency: the spoken follow-up question must be formed while listening, the dual-task that makes real consultations tiring, and only spoken practice loads it. And consultation flow: openings, signposting, transitions and closures have spoken rhythms, and rehearsing them aloud, platforms in the Geeky Medics real-time-voice mould being the accessible examples, builds the automaticity that frees attention for the actual listening. The same-scenario comparison is worth running once deliberately: one case in text, the same mutated case in voice, and the skills gap between your two performances is your personal answer to what voice adds.
What voice costs, and who it costs most
The access column, stated with the seriousness it deserves, because speech interfaces inherit speech recognition's documented inequities. Accent robustness: systems perform unevenly across regional and international accents, and the students most likely to be mis-transcribed include exactly the international candidates for whom rehearsal access matters most, the same lab-to-local gap the NHS phone-system episode illustrated at scale: /blog/yorkshire-accent-problem-nhs-ai-built-for-uk. Speech differences and disability: stammers, dysarthria, deaf and hard-of-hearing students meet a barrier text never posed, and a practice tool that excludes them has redistributed access, not expanded it. Environment: voice needs quiet and privacy, which shared accommodation and hospital libraries ration, while text practises anywhere. The tests to run before relying on any voice platform: your own accent and pace, transcribed accurately or not, in your real practice environment; the fallback question, whether text mode exists at parity; and, for programmes procuring centrally, accessibility testing with the students who will actually struggle, in procurement rather than in the complaints log.
The division of labour
A practical fortnightly pattern that takes both columns seriously. Text mode for content and structure work: new case types, differential coverage, question-set building, and all practice in noisy or shared environments, where text's privacy and accessibility are features. Voice mode for physics work: timed full consultations, interruption and flow rehearsal, and the pre-OSCE fortnight, when spoken automaticity pays directly. Both modes feeding the same downstream loop: misses converted to knowledge gaps, targeted questions, spaced retesting, per /blog/virtual-patient-to-qbank-close-the-loop-osce, because the modality changes what gets rehearsed, not what learning requires afterwards. And human calibration above both: the observed performance at meaningful intervals remains the validity anchor, per the triangulation model at /blog/ai-osce-feedback-vs-human-examiner, with voice practice feeding it rather than replacing it. Students who assign the modes by function get the consultation's physics and its content trained in parallel; students who pick one mode by novelty train half the skill and call it preparation.
Frequently asked questions
Does voice practice help with examiner-facing speech too?
Yes, presentations and viva answers share the physics: rehearsing case summaries aloud to a timer, with or without an AI listener, transfers directly, and it is the cheapest voice practice available.
My accent gets mis-transcribed; should I persist anyway?
Persist for your own fluency, since the speaking practice retains value even when the machine mishears, and weight the platform's feedback accordingly; report the failures too, because vendor accent performance improves on exactly that signal.
Is voice worth extra cost over text-only platforms?
Approaching an OSCE with flow and timing as your weak points, often yes; with content coverage as the gap, no, the division of labour answers the pricing question case by case.
Does practising with an AI voice change how I speak to real patients?
Watch for register drift: machine-directed speech runs flatter and faster, and the human calibration points exist partly to catch it; deliberately warming your opening and closes in voice practice inoculates most of it.
Is voice practice useful for non-OSCE assessments?
For any spoken performance: case presentations, handover, viva defence all share the timing-and-flow physics, and the same fortnightly division of labour covers them at no extra cost.