Transcription used to be a real line item. Human services charged $1.50 a minute or more, and you waited two or three days. Today AI transcribes an hour of audio in about two minutes for pennies, and if you run it on your own computer, the recording never leaves your possession. That last part is the piece most lawyers miss, and it’s the piece that matters most for privileged material.
What changed
OpenAI released a speech model called Whisper in 2022 and open-sourced it. That one decision collapsed the transcription market. Before Whisper, decent machine transcription cost real money and mediocre quality. After it, anyone could run near-human transcription for free, and a wave of tools got built on top of it.
The math today: a one-hour client call costs about $0.36 through the Whisper API. The same hour through a human service is $90 to $120. My firms transcribe hundreds of hours a month. At the old prices we simply wouldn’t have done it, which means all that information would have stayed locked in audio files nobody ever listened to twice.
The confidentiality angle nobody talks about
Every cloud transcription vendor will show you a security page with the usual certifications. Fine. But there is a cleaner answer for privileged audio, and it costs about $40 once.
On-device transcription means the Whisper model runs on your own Mac. You drag an audio file in, the processing happens on your laptop’s chip, and the transcript appears. No upload. No vendor. No terms of service governing what happens to a recording of your client describing their case. The audio never touches a network.
For a solo lawyer or a small firm, this ends the vendor diligence question entirely. There is no third party to vet because there is no third party. MacWhisper is the easy version of this. whisper.cpp is the free command-line version for anyone comfortable in a terminal. Both run the same underlying model the cloud services use.
I tell lawyers in my communities to sort their audio into two buckets. Privileged and sensitive material goes local. Everything else can go to the cloud for speed and convenience.
The tools compared
| Tool | Price | Where audio goes | Best for |
|---|---|---|---|
| MacWhisper | ~$30-60 one time | Stays on your Mac | Privileged audio, no subscription |
| whisper.cpp | Free | Stays on your computer | Same, if you’ll use a terminal |
| Whisper API | ~$0.006/min | OpenAI cloud | Bulk jobs, automation pipelines |
| Rev AI | ~$0.25/min | Rev cloud | Speaker labels, legal vocabulary |
| Rev human | $1.99/min | Rev cloud + human | Highest accuracy short of certified |
| HappyScribe | ~$17/mo | Their cloud | Subtitles, many languages |
| Descript | ~$16-24/mo | Their cloud | Editing audio and video by editing text |
If you only get one, get MacWhisper. It covers the sensitive work, and the cloud options are always there when you need speed on routine audio.
What lawyers actually transcribe
Client call recordings are the obvious one. At my lemon law firm, intake and doc-collection calls get transcribed so nobody has to re-listen to a 40-minute call to find one detail about a repair visit.
Voicemail triage is the sleeper use. Transcribe every voicemail automatically and staff can scan 30 messages in the time it took to listen to 4. Urgent ones surface immediately.
Depo prep sessions and witness interviews, transcribed same-day, so you review the record while it’s fresh instead of relying on notes. Dictated memos, if you think out loud better than you type, and there’s a whole dictation workflow built around exactly that.
And Spanish-language calls. Whisper handles Spanish about as well as English, which is a big deal for consumer practices. A meaningful share of my firm’s clients prefer Spanish, and transcription plus translation means an English-speaking case manager can review any call without waiting on a bilingual colleague.
What not to use it for
Official deposition and court transcripts. Full stop. Certified court reporters remain the standard, courts require the certification, and no AI transcript substitutes for it. Use AI around the deposition, never as the deposition record. Anyone selling you otherwise is selling you a sanctions motion.
Accuracy and speaker labels
Expect 92 to 97 percent accuracy on clean audio. Legal terms of art usually survive; proper names and case citations often don’t, so search the transcript for names and fix them. Bad phone audio, heavy accents, and people talking over each other drag accuracy down fast.
Speaker labels (who said what) are where local tools lag. MacWhisper’s higher tier handles diarization, but Rev AI and Descript do it better. For a two-person client call you rarely need labels. For a four-person strategy call you do.
One legal reminder before you record
Recording consent laws still apply. Eleven states require all-party consent, California and Florida among them. Transcribing a recording you had no right to make doesn’t clean it up. The fix is a standard disclosure at the top of every recorded call. Thirty words, problem solved, and it doubles as good client communication.
If most of your recordings come from meetings rather than phone calls, an AI notetaker handles the consent banner and the recording in one step.
The workflow: transcript to action items
A raw transcript is an intermediate product. The value shows up when you feed it to Claude with a prompt like: “Summarize this client call in five bullets, list every action item with an owner, and flag anything that sounds like a deadline.” That turns a 40-minute call into a 30-second read, and it works on every call, every day, without anyone taking notes.
At my firms this pair, transcription plus summarization, quietly became one of the highest-return automations we run. Nobody types call notes anymore.
Set up local transcription in 10 minutes
- Buy MacWhisper from the developer’s site (Gumroad) or grab the free tier to test. The Pro version runs $30 to $60 depending on current pricing.
- Open it and download a model. Start with “large-v3-turbo” if your Mac has Apple silicon, “medium” if it’s older.
- Drag in a test audio file. A voicemail or a memo you dictate on your phone works fine.
- Check the output, then set your default export format to text or markdown.
- Turn off any update-check or telemetry options if you want the machine fully quiet, and confirm in the settings that you’re running a local model.
- Transcribe one real client call this week and paste the transcript into Claude for a summary.
Step 6 is the one that counts. Do it once and you’ll stop paying anyone $1.50 a minute for anything short of a certified transcript.