Conversation Capture: audio that survives the hospital.
An AI scribe is only as good as what it hears. This is the story of the capture layer we built for a clinical documentation product: recording a live patient encounter on a clinician’s phone and getting it to the cloud reliably, even on hospital Wi-Fi that drops mid-sentence.
Listening to the world, in real time
A clinical encounter doesn’t pause for the software. If a scribe is going to draft the note, the conversation has to be captured continuously and completely: low latency, nothing dropped, resilient to the dead zones every hospital has. Miss thirty seconds of a risk assessment and the note is wrong in a way that matters.
Plenty of tools can record audio on a phone. Very few can stream a live conversation to the cloud for the length of a full clinical session without losing data when the network misbehaves, the app is backgrounded, or the device runs low on resources. That gap is what this project closed.
What conversation capture means
Passively capturing audio from the room and moving it, continuously and safely, to where it becomes useful: a drafted note, a structured summary, a decision-support signal. Unlike simple recording, capture is a flow. The audio is always in motion, and the system's job is to make sure none of it goes missing along the way.
Why this is hard on a phone
Four realities of mobile clinical environments shaped the whole design.
Continuous capture
A session can run over an hour. Capture has to stay live for all of it, without gaps.
Unreliable networks
Hospital Wi-Fi and cellular handoffs drop constantly. The recording cannot depend on a stable connection.
Real-time expectations
The note should be drafting while the visit happens, not uploading after it ends.
Device constraints
Phones background apps, reclaim memory and throttle work. The pipeline has to be a good citizen and still never lose audio.
Design so nothing is ever lost
Instead of streaming one long, fragile recording, the system breaks the conversation into small pieces as it happens. Each piece is secured on the device the moment it exists, then shipped to the cloud independently. If any piece fails to arrive, only that piece is retried. The visit itself is never at risk.
Delivery is aware of the world around it. When the network disappears, uploading pauses on its own; when it returns, delivery resumes exactly where it left off, even across app restarts. Progress is tracked piece by piece, so the clinician can see the session moving to the cloud in real time rather than trusting a spinner.
Resilient by default
Interruptions are the expected case, not the exception. Recovery is automatic and invisible.
Light on the device
Long sessions never pile up in memory, so the app stays responsive for the whole encounter.
Ordered and complete
The cloud receives the conversation exactly as it happened: in order, with nothing missing.
What the clinician experiences
All of that machinery exists so the person in the room can ignore it. Start the session, see it's listening, and give the patient full attention.
Confidence at a glance
- A live waveform confirms the room is being heard, without stealing attention.
- Pause and resume respect the moments that shouldn’t be recorded.
- Sessions have sensible safety limits and warnings, so nothing runs on unnoticed.
- If the phone loses signal, the session simply continues. There is nothing for the clinician to manage.
Built-in safety limits warn before a session runs long. Pause and resume preserve session state; silent gaps are detected automatically.
What it delivered
No lost encounters
Complete recordings even through network drops, app backgrounding and long sessions.
Notes ready sooner
Audio reaches the cloud during the visit, so drafting starts before the clinician leaves the room.
Works on real devices
Dependable across both major mobile platforms and the mixed device fleet of a real clinic.
A reusable foundation
The same capture-and-deliver layer now underpins other real-time features of the product.
Toward a smarter clinical audio layer
Once capture is dependable, it becomes a foundation for clinical audio intelligence: the same layer that powers real-time note drafting can carry richer capabilities.
Ambient understanding
Recognize the shape of a visit as it unfolds, not just the words.
In-the-moment prompts
Surface gentle, assistive cues while the encounter is still happening.
Two-way audio
Extend the same reliability to telehealth and remote care conversations.
Deeper decision support
Feed downstream intelligence with a complete, trustworthy record of the visit.
Reliability is the feature
The hardest part of conversation capture isn’t recording audio. It’s guaranteeing that a live, unrepeatable conversation survives everything a real clinical environment throws at it. By designing for failure from the first day, we shipped a capture layer clinicians never have to think about, which is exactly the point. If you’re building a product where real-time data has to arrive complete, these same principles apply.