Can you dictate SOAP notes from an Apple Watch? What actually works for ambient capture
Short answer: yes, you can dictate into an Apple Watch and get a SOAP note out the other end — but only as dictation, with the clinician deliberately speaking to the device after the encounter. As an ambient capture device that quietly records a live consultation and turns it into a structured note, a smartwatch is the weakest option available to you, and the reasons are acoustic rather than a matter of software.
That distinction is the whole question, and it gets skipped constantly. Dictation and ambient capture put completely different demands on the microphone, and everything downstream inherits whatever the microphone gave you.
Dictation and ambient capture are different problems
Dictation is a single, cooperative speaker talking directly at a microphone from a short distance, in a quiet room, using deliberate phrasing. That is close to the easiest possible input for a speech model. Word error rates in the low single digits are routine, and a watch handles it perfectly well — the mic is near your mouth and you are speaking to it.
Ambient capture is a two-or-three-speaker conversation at conversational volume, across a room, over air conditioning and corridor noise, with interruptions, overlapping speech and a patient who may be facing away. The signal reaching the microphone is far weaker relative to the noise, and the system additionally has to work out who said what.
A smartwatch fails the second case for a structural reason: the microphone is small, it sits at your wrist, and your wrist moves. Every arm movement changes the distance and angle to both speakers, and fabric brushing the case produces broadband noise directly on the mic element. Wrist-worn microphones are also tuned aggressively for near-field voice — good for talking into your watch, actively unhelpful for picking up someone across a desk.
What the pipeline needs to produce a usable note
The reason capture quality matters so much is that errors compound down the chain, and the chain is longer than most people expect.
Raw audio has to be segmented into speech regions, then attributed to speakers, then transcribed, then that transcript has to be turned into a structured note — subjective, objective, assessment, plan — with clinical entities extracted and mapped to the fields your EHR expects.
Speaker attribution is the stage that quietly decides whether the note is usable. If the system cannot reliably separate clinician from patient, the note becomes dangerous rather than merely wrong: symptoms the patient described get attributed to the clinician’s assessment, or the clinician’s differential gets recorded as the patient’s reported history. Diarisation degrades sharply with distance and noise, which is exactly what a wrist microphone gives it.
Transcription errors then propagate. A drug name misheard as a similar-sounding one survives into the structured note unless something downstream validates it, and a model summarising a transcript has no way to know a word was wrong. We covered how that structuring stage works end to end in building an AI medical scribe, and what the output actually looks like in a real audio-to-SOAP example.
Producing a note a clinician will sign is the whole job, and it is where most of the work in ambient scribe development sits.
Capture setups, ranked by what survives a real room
A phone on the desk, screen down, is the practical default. Modern phones have multiple microphones and run beamforming across them, which is genuinely effective at picking a speaker out of room noise. It is stationary, it is already in the clinician’s pocket, it needs no procurement, and it consistently outperforms every wearable we have tested. If you do one thing, do this.
A dedicated USB or Bluetooth conference microphone is the quality ceiling. Omnidirectional, positioned between clinician and patient, designed for exactly this acoustic problem. Worth it in a fixed consulting room; impractical for ward rounds or home visits.
A wired lapel mic works for dictation-heavy workflows where the clinician narrates findings during or immediately after the encounter. Close to the mouth, stable distance, excellent signal — but it captures the wearer far better than anyone else, so it is a dictation tool, not an ambient one.
A smartwatch is last for ambient capture, and fine for dictation. If the workflow is “clinician steps out and speaks a two-minute summary into their wrist”, a watch is genuinely convenient and the accuracy will be good. If the workflow is “record the consultation”, use the phone that is already in the same pocket.
The constraints that decide it in practice
Battery and thermal limits are the ones that surprise people. Continuous audio capture with the screen off is manageable on a phone across a clinic day; on a watch it competes with everything else the watch is doing, and sustained recording plus streaming upload is one of the more expensive things you can ask of it. A capture device that dies at 3pm is not a capture device.
Connectivity shapes the architecture more than the device does. If audio streams to a server for transcription, a dropped connection mid-consultation must not lose the encounter — that means buffering locally and reconciling on reconnect, which is real engineering rather than a setting. On-device transcription avoids it but constrains model size, and small models degrade fastest on exactly the noisy multi-speaker audio a wearable produces.
And storage on the device is PHI the moment it exists. A buffered recording sitting in a cache on a personal watch is protected health information on a consumer device, inside your compliance boundary, with your retention obligations attached. That is a policy question before it is a technical one, and it is covered in more depth in HIPAA-compliant AI architecture.
Consent is a workflow requirement, not a legal footnote
Ambient recording of a clinical encounter requires the patient to know it is happening. Beyond the legal position, which varies by jurisdiction, it changes the system design: there has to be a visible, unambiguous indication that recording is active, a way to stop it mid-encounter, and a way to discard what was captured. A capture device with no screen and no clear indicator makes that materially harder to do honestly — another point against the wrist.
What we would build
For most clinics: phone as capture device, buffered locally, streamed to a transcription pipeline with diarisation, structured into a draft note with per-field confidence, surfaced to the clinician for review before anything enters the record. The watch, if it appears at all, is a control surface — start, stop, flag a moment — rather than a microphone.
That review step is not a compromise. A system that drafts for a clinician to approve sits in a very different regulatory position from one that writes to the record autonomously, and it is also the only reliable way to catch the transcription errors that survive the pipeline. Our AI medical scribe case study auto-fills more than ten clinical forms from a single consultation and still routes every one of them through clinician review, for exactly that reason.
Tools referenced
- Deepgram — streaming speech-to-text with speaker diarisation
- WhisperX — alignment and diarisation on top of Whisper
- pyannote.audio — speaker diarisation toolkit
- Apple AVAudioSession — capture configuration and interruption handling on iOS and watchOS
EpochC builds speech and audio ML systems and AI medical scribes that run in production clinics. If you are choosing a capture setup for a clinical documentation build, book a technical discovery call — bring a sample recording from the room you actually work in and we will tell you what your pipeline will do with it before you commit to hardware.
Related: ambient clinical intelligence explained · building an AI medical scribe · AI clinical documentation