For most of Terra's pilot, the transcript arrived after the visit. A clinician recorded the conversation, uploaded the file, waited for transcription, reviewed it, and then prepared the draft. It worked. It also meant the most important check, "did it hear that right?", happened after the patient had gone home.
Today's release changes that. Terra now transcribes the visit live, in the Transcript tab, while you talk. This post explains why we built it this way and what happens behind the screen.
The transcript belongs in the room
Clinicians told us the same thing in different words: if I can see it while we are talking, I can fix it while we are talking. A medication name heard wrong, a dose the patient corrected halfway through a sentence, a symptom they described differently the second time. With a live transcript, you notice in the moment and ask. With a transcript that arrives later, you guess, or you go back to the recording.
There is a second benefit. When the transcript is already on screen, the draft can be ready almost as soon as the visit ends. Stop the session, confirm the speakers, append the text, and choose Prepare draft. The note is waiting before the next patient is roomed.
Consent comes first, every time
Recording a clinical conversation is a decision the patient shares. So the first thing in the live panel is not a microphone button. It is a checkbox: I confirm all parties consented to recording this visit. Until it is ticked, the start button does nothing.
We thought hard about whether that is too much friction for a clinician who does twenty visits a day. We decided it is the right amount. It takes a second, it makes the moment explicit, and it gives the practice a clear point in the workflow where consent is confirmed. It does not replace your state's rules or your practice policy, which may require more, but it means Terra never starts listening by accident.
Terra also asks for microphone permission before it opens a billed session. If the browser refuses the microphone, nothing is reserved and nothing is charged.
What happens to the audio
Audio streams from your browser straight to our speech provider over an encrypted connection, using a token that works for a single session and expires in 60 seconds if it is not used. The token is never stored. Terra's servers never receive the live audio, and Terra does not keep a copy.
What Terra does keep is the text. As each speaker finishes a turn, the finalized words, their timestamps and a neutral speaker label are saved to a journal on the visit every couple of seconds. If your browser crashes or your laptop goes to sleep, the turns already finalized are still there when you come back. A stored turn can never be rewritten; your edits happen later, in review.
To help with clinical vocabulary, Terra sends the speech model a short list of key terms: the patient's approved, active medications and problems first, then a small static list of clinical words. That is the only chart information involved, and it helps the model hear "lisinopril" rather than something that sounds like it.
Grey words, final words
While someone is speaking, the words appear in grey. They are the model's current best guess and they can still change. When the speaker finishes the turn, the text is formatted and turns final. This mirrors what the model is actually doing, and it stops you from correcting a word that is about to correct itself.
Speaker labels you confirm
The speech model separates voices and labels them neutrally: A, B, C. It does not know who is the clinician, who is the patient and who is the daughter who came along to help. So when you stop the session, Terra asks you.
Each speaker gets a role you choose: Clinician, Patient, Interpreter, Caregiver, Other or Unknown. To save time, the first speaker starts as Clinician and the second as Patient, which is right most of the time. When it is not, change it. Then read the role-labelled text, correct anything wrong, confirm your review, and choose Append to visit transcript or Replace visit transcript.
Nothing reaches the note until you do that. If the session was a false start, Discard live text closes it without changing the visit.
We made the review step deliberate because the transcript is evidence. Every sentence in the draft cites the transcript turn it came from. If the speaker roles are wrong, the draft might attribute a patient's worry to the clinician's assessment. A ten-second check prevents that.
Dictation into one section
The same panel handles dictation. The microphone next to the Patient context field, and next to each note section heading, starts a session for just that field. Speak your assessment or your plan, stop, review, and choose Insert at cursor. It is saved like any other edit.
Billing by the session, settled to the second
Our speech provider bills streaming by session time, not by how much audio is sent. So Terra does the same. When you start a session you choose a session limit. Terra reserves credits for that limit up front, at 12 credits per minute. When you stop, the provider confirms how long the session lasted, Terra checks that against its own clock, and you are charged for the actual time, rounded to whole seconds. The rest of the reservation is returned straight away.
Two cases are handled conservatively:
- If the connection drops and the provider never confirms the end of the session, Terra cannot know exactly what was billed. The reservation is held, and a practice manager reconciles it. It is never charged twice.
- If a session is left running and the browser never ends it, Terra settles it after the session limit at the reserved amount. That is a known upper bound, because the provider cannot bill past the limit.
If the socket drops mid-visit, Terra reconnects once and continues numbering turns from where it left off, so applying the latest session applies the whole visit.
What it is not
Live transcription is not a recording you keep, and it is not a replacement for listening. It does not diagnose, summarize or decide anything. It is a faster, more checkable way to get the words of the visit into the chart, with the clinician confirming every step.
Live transcription is available on every plan. If you have questions about how it fits your consent process, write to us at hello@useterra.si. The Help Center guide walks through each step.
