Legal transcription software: where automatic speech recognition stops
Legal transcription software gives you a searchable draft of a deposition or hearing, not a certified transcript. Here is where the line sits.
Automatic transcription is good for one half of legal work and wrong for the other. It will give you a searchable draft of a deposition, a hearing recording or a client call in minutes, which is how you find the moment something was said without scrubbing through three hours of audio. It will not give you a certified transcript, because certification generally means a person signing an attestation, and no software signs anything. With legal transcription software the useful question is where that line falls.
What the automatic draft is genuinely good for
Four jobs, and they are the ones that eat a paralegal’s week.
Finding the moment something was said. A draft with timestamps turns three hours of audio into a text file you can search. Type the witness’s phrase, get a time, open the recording there. The words in the draft do not have to be perfect to point you at 01:14:20.
Triage. Ten hours of recorded calls usually contain ten minutes that matter. Skimming a rough transcript is far faster than listening at 1.5x.
Working drafts for internal use. A chronology, a witness summary, a memo to the file, a list of questions for the next session. None of these leave the firm, none of them are filed, and a draft with a few wrong words in it does the job as long as you check the parts you quote.
A starting point for a human transcriber. The one that saves real money, covered below.
What these have in common is that a mistake costs you a minute, not a case. That is the test. If a wrong word in the transcript can only send you back to the audio, automatic is fine. If a wrong word can end up in a filing, a quotation or a record, it is not.
A certified transcript is a signed human product
Certification generally works like this: a named person states that the transcript is a true and accurate record of what was said, and puts their name to it. That is a human act with a human on the hook for it. A model produces text with no author and nobody answerable for it, so there is nothing to attest to.
Requirements differ by jurisdiction, by court and sometimes by agreement between the parties, covering who may certify and what form the certificate takes. We have not verified any of them, so we are not summarising them here. Check the rules that apply to your matter before you assume any transcript, automatic or human, is acceptable.
The practical version is usually simple. Anything that is going to be filed, served, quoted in a pleading, put to a witness, or relied on as a record goes to a human. Everything upstream of that can be automatic.
Speaker attribution is where a deposition lives, and the model does not do it
In a deposition the identity of the speaker is half the meaning. “I saw him leave at nine” is a different document depending on who said it, and a transcript that runs the examiner and the witness together as one block of prose is close to useless as a record.
Speech recognition models like OpenAI’s open-source Whisper do not separate speakers. Diarisation is a different task done by different software. The Whisper authors went further and trained the behaviour out deliberately. Their paper says they saw the models “transcribe plausible but almost always incorrect guesses for the names of speakers”, because transcripts in the pre-training data often name the speaker, which encourages the model to guess, and the last 30 seconds of audio rarely tell it who that is (section 2.4). Their fix was to fine-tune on transcripts without speaker annotations, which “removes this behavior”.
So a clean automatic transcript of a four-person examination arrives as one undivided block. You can add the labels yourself from the timestamps, which our post on two-speaker interviews walks through, and for two voices it takes a few minutes. For four voices with interruptions, it is slow, and it is exactly the kind of slow work a transcriber does better.
Crosstalk makes it worse. When two people speak at once the model writes one stream of words, usually the louder one, and the interruption disappears. In a heated examination the overlap is often the part you care about.
Names, case numbers and figures are the tokens a small model gets wrong
The errors a speech model makes are not spread evenly across a transcript. Common words are safe. The tokens that carry legal weight are the fragile ones: surnames, company names, street names, case numbers, dates, dollar amounts, statute and section numbers, medical and technical terms. They are rare in the training data, they often have no strong context around them, and the model falls back on whatever sounds similar and is more common.
A base model, which is what a browser tool can realistically download and run, is weaker at this than the large ones. Courtroom and conference-room acoustics make it weaker still: a far-off microphone, a hard-surfaced room, a speakerphone on a table, a witness turning away mid-answer. Our post on the seven mistakes to fix in a transcript covers how to hunt these down, and numbers and names are the first two for a reason.
In this setting, “read it before relying on it” is not a disclaimer. Every figure and every proper noun in a passage you intend to use has to be checked against the audio.
Fluent invented text is a real risk in a legal file
Speech models can write text over silence. The Whisper paper lists, among the failure modes it still has, “getting stuck in repeat loops, not transcribing the first or last few words of an audio segment, or complete hallucination where the model will output a transcript entirely unrelated to the actual audio”. In a podcast transcript that is annoying. In a deposition draft it is a sentence that reads like testimony and was never spoken.
Two shapes are easy to miss. A plausible single sentence over a pause, which looks like every other line around it. And the clipped opening or ending of a segment, where the first words of an answer are gone and what remains still parses as a complete answer. We explain the mechanism and the defences in why speech recognition invents words in silence. The rule for legal work is that no passage goes anywhere from the draft without being heard.
Where legal transcription software actually saves money
The saving does not come from replacing the transcriber. It comes from shrinking what you send them.
- Transcribe everything locally. Every deposition recording, every hearing, every recorded call, as soon as you have the file. Cost is time, not money.
- Search and triage the drafts. Find the passages that matter. On ten hours of recordings, this is usually under an hour.
- Mark the passages. Note the timestamps of the parts you will quote, file, or put to a witness. Export SRT if you want the timings visible next to the text.
- Send only those passages out. Clip the audio to the marked ranges and send the clips for human transcription. You pay per minute of audio, so twenty minutes instead of ten hours is where the money goes.
- Keep the draft internal. Use it for the chronology, the summary and the memo. Do not let it out of the building.
| Task | Automatic draft | Human transcript |
|---|---|---|
| Finding when something was said | Yes, with timestamps | Overkill |
| Triaging hours of recordings | Yes, this is the main use | Overkill |
| Chronology, witness summary, internal memo | Yes, check anything you quote | Not needed |
| Who said which words | No, one undivided block | Yes, this is the core of the job |
| Overlapping speech and crosstalk | No, one speaker survives | Yes |
| Exact names, figures, citations | Check every one against the audio | Yes |
| Anything filed, served or quoted | No | Yes, with the attestation |
One more reason to do step 1 locally: a deposition recording is somebody else’s confidential material, and a tool that runs in your browser never sends the file anywhere. FreeTranscribe does that, free and with no length cap, on desktop Chrome or Edge with a working graphics card, in English, using the base model. It is a draft generator. Treat it as one.
Frequently asked questions
Can an automatic transcript be certified if a person reads it first? Who may certify a transcript, and what that certification has to say, varies by jurisdiction and forum. A person correcting a machine draft against the audio is a normal way to produce a transcript, but whether the result is acceptable in your matter is a question for the rules that apply to it, not for us.
Does a larger model solve the names and numbers problem? It reduces it. The Whisper authors report that larger models made “steady and reliable progress on reducing perception-related errors such as confusing similar-sounding words”, while the long-form errors “seem more stubborn in nature”. Better audio helps more than a bigger model: a microphone near each speaker changes everything.
Is it safe to run client recordings through a browser tool? It depends entirely on where the audio goes. A tool that processes the file in your browser does not upload it. A cloud service does, which generally makes it a processor handling your client’s data, with the contractual and confidentiality questions that follow.
How long does the human pass take if I do it myself? Budget roughly an hour per hour of clean two-speaker audio to correct a draft properly, more for multi-speaker or poor recordings. That is the number to compare against a quoted per-minute rate when you decide what to send out.
This post is general information, not legal advice.
Sources, checked 15 September 2026
- https://arxiv.org/html/2212.04356 - Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision”, section 2.4 (incorrect guesses for speaker names, fine-tuned out) and section 6 (repeat loops, clipped segments, complete hallucination; larger models and perception errors)