Engineering

Real-time speech-to-text in the courtroom: from microphone to record

By the exma team · July 23, 2026 · 7 min read

Why real-time matters in a courtroom

A transcript delivered two weeks after the hearing is an archive. A transcript that exists while the witness is still on the stand is a working tool:

• Instant read-backs. "Could you repeat the witness's last answer?" becomes a glance instead of a pause in proceedings.
• Live search. Counsel can search the morning's testimony during the afternoon session — before cross-examination, not after.
• Accessibility. Deaf and hard-of-hearing participants follow the proceeding as it happens, not from a delayed summary.
• Same-day drafts. Prep for day two starts the evening of day one.

Getting there means solving a harder problem than transcribing a finished file: the system must be accurate and fast, on audio it hears exactly once, in rooms that were not designed for microphones.

The pipeline at a glance

Stage 1: Capture — no bot in the room

Consumer note-takers capture audio by joining a video call as a participant. That fails immediately in legal settings: courtrooms and depositions aren't Zoom calls, and even when a hearing is remote, a visible bot participant is often unacceptable.

exma captures at the source instead: the microphone for the room, the computer's system audio for anything playing through it — a remote hearing on any platform, a phone call, played-back evidence — or both at once for hybrid proceedings. Capture quality sets the ceiling for everything downstream: a microphone near each speaking position beats an expensive microphone across the room, and digital system-audio capture beats re-recording a speaker's output every time.

Stage 2: Streaming recognition — words in milliseconds

Batch transcription reads a whole file, with full context, before producing output. Streaming recognition can't wait — audio arrives in small chunks and text must appear almost immediately. The engine handles this with interim and final results: a provisional hypothesis appears within a fraction of a second, then refines itself as more context arrives, and locks in as final once the phrase is settled.

Watch a live transcript closely and you'll see it: the last few words shimmer and self-correct as the sentence completes. "Their motion" becomes "the remotion" becomes "the remote deposition" as the acoustic evidence accumulates. That behavior isn't a glitch — it's the visible half of how streaming models trade a moment of provisional text for low latency without sacrificing final accuracy.

Stage 3: Speaker diarization — who said what

A courtroom transcript without speaker attribution is nearly worthless: testimony must be attributable to be citable. Diarization separates the audio into speakers by voice characteristics, in real time, so the draft reads as a dialogue rather than a wall of text. Labels start generic — Speaker 1, Speaker 2 — and get named once (THE COURT, THE WITNESS, MR. CHEN), applying across the whole transcript.

Diarization is also where courtroom audio is hardest. Legal proceedings feature overlapping speech — objections exist to interrupt — plus similar-sounding voices and varying distances from microphones. Good capture discipline (stage 1) pays for itself here: clean, well-separated input keeps speaker boundaries sharp when it matters most.

Stage 4: Legal vocabulary — voir dire, not "vwa deer"

General-purpose recognition models are trained on general speech, and it shows the moment testimony turns technical: voir dire, res judicata, subpoena duces tecum, drug names in an expert's testimony, case citations. A domain-tuned engine weights recognition toward legal terminology, Latin phrases, and procedural language, so specialized terms land correctly the first time — because in a verbatim record, "guessing something phonetically similar" is an error, not a rounding.

Stage 5: Timestamps and structure — text that stays tied to audio

As recognition and diarization stream in, the pipeline assembles the working record: punctuation and capitalization, speaker-labeled turns, and timestamps that sync every passage to the recording. That sync is what makes the draft verifiable — any disputed line can be checked against the audio in one click, which is also what makes [inaudible] passages honest: they carry a timestamp pointing at exactly what to re-listen to.

In exma: the live draft is searchable while the proceeding is still running, and exports to TXT, DOCX, and PDF with speakers and timestamps intact. Uploaded audio and video follow the same pipeline in batch, minus the latency constraint.

Stage 6: From live draft to official record

The live transcript is a working draft — immediately useful for read-backs, search, and prep. Becoming the official record takes the steps machines don't do: review against the audio, resolution of unclear passages, confirmation of speaker identity, and a signed certification. exma delivers certified versions reviewed before delivery for exactly this reason: the pipeline's job is to make the accurate draft nearly instant, so human effort goes to judgment instead of keystrokes.

What the courts require of that final artifact — authentication, certificates, chain of custody, formatting — is its own topic, covered in What makes a transcript court-admissible?

The part you don't see: security

Courtroom audio is some of the most sensitive data an organization handles, so the pipeline's invisible properties matter as much as its visible ones. In exma: audio and transcripts are encrypted in transit and at rest, chain-of-custody tracking runs from capture to delivery, access is controlled and logged under SOC 2-aligned controls, and customer audio and transcripts are never used to train AI models. A transcription pipeline that quietly turns testimony into training data has no place in legal work.

Frequently asked questions

Is real-time transcription accurate enough for the official record?

The live stream produces a highly accurate, timestamped working draft. The official record comes after review against the audio and certification. Real-time gets you a usable transcript in seconds; review makes it official.

Does real-time speech-to-text replace court reporters?

It changes the work rather than removing the role: the machine captures every word in real time, while humans resolve unclear audio, confirm speakers, and certify. Teams use live AI transcription where a stenographer isn't available or budgeted, and human review wherever certification is required.

What audio setup produces the best live transcription?

Placement over price: a microphone near each speaking position, minimal room echo and background noise, and digital system-audio capture for remote participants.

Does exma join court proceedings as a bot?

No. exma never joins anything as a participant — it captures microphone, system audio, or both. No visible participant, no meeting-platform dependency.