Hearing transcription that holds up in court: a quiet revolution for Indian litigation
Indian courts produce thousands of hours of audio every week. Most of it sits unused because transcription is too slow. Here is how AI changes that, what holds up to scrutiny, and what does not.
TL;DR
A 90-minute hearing used to take an associate a full day to transcribe, and the firm usually skipped it. Now the same 90 minutes comes back as a speaker-tagged, verified transcript in 12 minutes. We've installed this at six Indian firms in the last quarter. The thing nobody tells you is that the transcript is not the win. The win is what the firm does with the searchable archive afterwards.
Why transcription has been broken at most firms
For 30 years, the calculus has been the same. A junior associate goes to court, takes notes, comes back, and writes up a memo. Two days later, after one round of partner review, the memo lands on the matter file. The audio recording, if it was made at all, sits on a phone or a hard drive. Nobody listens to it again.
The friction is the transcription. A 90-minute hearing takes a competent associate roughly six hours to transcribe verbatim. That cost is not justifiable for most matters. So the firm settles for the memo and accepts that the verbatim record is gone.
That trade-off was rational when transcription was a six-hour job. It stops being rational when transcription is a 12-minute job.
What we install
The workflow takes audio in (a phone recording, a court microphone feed, a Zoom file from an arbitration hearing) and returns four things.
A verbatim transcript with speaker labels. The labels are auto-detected and then a human (usually the associate who was in the hearing) corrects the speaker identification in about two minutes. The transcript itself is good enough that the associate is editing names, not text.
A timestamped index. Every paragraph has a time stamp linked to the audio. If a partner wants to hear what the witness actually said at 14:32 about the bank statement, one click takes her to the audio at that exact second.
A chronology update. If the matter has a chronology document on file, the workflow proposes updates: new dates, new admissions, new contradictions to flag. The associate accepts or rejects each one.
A contradictions table. This is the one nobody asks for in advance and nobody gives up after the first month. The workflow scans the new transcript against the prior testimony of the same witness, and flags every place the witness has said something inconsistent. It is not a finished cross-examination. It is a starting point that used to take a senior associate a full day to build by hand.
The native-language piece
Indian litigation is multilingual. Hearings in Mumbai courts happen in English, but witness depositions outside courtrooms happen in Marathi, Gujarati, Hindi. Cross-border arbitrations have witnesses in three languages. The transcript workflow handles this.
For a Marathi deposition, you get a side-by-side transcript: Marathi on the left, English translation on the right. Both have time stamps. A bilingual associate verifies the translation before it goes on file. The translation is treated as a starting point, not a final document, because bar council rules require a human-verified translation for anything that goes into evidence.
What this means in practice: the firm stops paying ₹15,000 per hearing to an external transcription vendor. It stops waiting two weeks for the transcript to come back. The senior associate does a 30-minute verification pass on the day of the hearing instead.
The thing that actually changes
The transcript itself is useful. The change is the archive.
After 90 days of running the workflow, the firm has a searchable archive of every hearing it has done. That archive is queryable. Six months in, when the same opposing counsel cites a case, the firm can pull up every prior hearing where that counsel made a similar argument, in seconds.
We have a client who used this last quarter to find a contradiction in opposing counsel's position that went back two years across three different matters. The senior partner had a vague memory of the previous statements. The archive turned the vague memory into a citation with a time stamp.
Nobody asks for that capability when they install the transcription workflow. Everybody refuses to give it up after they have it.
What we do not do
We do not push transcripts to court without verification. We do not treat the AI output as the final document. We do not auto-generate hearing memos and bypass associates. The transcript is a tool that turns six hours of associate time into 30 minutes of verification time. The associate is still in the loop, and still responsible for what goes on file.
We also do not do real-time transcription in court. The technology exists, but Indian courtroom acoustics are difficult, and we have not seen a deployment that holds up to scrutiny. Post-hearing, on a clean recording, we trust the workflow. In the courtroom, in real time, with three people talking at once, we don't.
If you want to see what 12 minutes of transcript looks like for one of your existing matters, send us a 5-minute sample audio file. Book a teardown and we will run it through the workflow on the call.
Frequently asked
Transcripts are not evidence in themselves. They are a working tool. The audio recording is the source. We treat the AI transcript as a senior associate's first pass that a human verifies before it goes on file. That is the same standard the firm would apply to a junior's transcription anyway. We have not had a single transcript challenged in court because nobody puts an unverified transcript in court.