Why Defenders Are Looking at Whisper in the First Place

The discovery dump lands in your inbox: bodycam from three officers, a two-hour interrogation video, jail calls, a 911 recording. Somewhere in those files are the two minutes that win the suppression hearing. Scrubbing through a three-hour video file to find them is the job nobody billed for — and it's why defenders keep searching for a faster way to get a searchable record.

OpenAI Whisper sits at the top of most search results for a reason. It's free, open-source under an MIT license, and downloadable from OpenAI's official repository — no subscription, no vendor lock-in. For a solo practitioner or an underfunded public-defender office, that price tag is hard to argue with. A study of AI transcription on bodycam footage captures the appeal plainly: AI transcription lets public defenders "find the two minutes they're looking for without having to scrub through a three-hour video file."

That's the promise. The reality requires knowing exactly what you're getting, because "Whisper" actually means three different things:

• Open-source Whisper — the MIT-licensed models (large-v3, turbo) you download and run locally or offline, from OpenAI's GitHub repository.
• The hosted whisper-1 API — no local setup, billed at $0.006 per minute ($0.36 per audio hour).
• Newer managed OpenAI models — gpt-4o-transcribe and gpt-4o-mini-transcribe, released March 2025, which reportedly have lower error rates but are not open-source and are not newer versions of open-source Whisper.

Each path gets you raw speech-to-text. None of them, on its own, gets you a review-ready evidence workflow — and the gap between "transcript" and "usable discovery review" is where most DIY projects stall. Interrogation footage and jail calls are multi-speaker recordings; Whisper, on its own, turns audio into text and doesn't tell you who spoke. Fixing that means bolting on separate tooling, and the accuracy numbers on real-world audio — 12.4% word error rate on ordinary consumer-mic speech versus the 3.6% headline benchmark — mean every output needs a human pass before it touches a case.

That trade-off is worth understanding before you install anything. The rest of this article walks through all three ways to actually get Whisper — the local install, the API, and what a managed alternative like Evidence Copilot handles for you — with the setup requirements, the real costs, and the failure modes that matter when the transcript ends up in front of a judge.

The Three Ways to Get Whisper — Local Install, Hosted API, Managed Models

There are three ways to "get" Whisper, and they're not the same product. The open-source model you install locally, the hosted API you pay per minute, and OpenAI's newer managed transcription models behave differently, cost differently, and carry different limits — so it's worth knowing which one you're actually dealing with before you point it at a jail call or a bodycam file.

Path 1: Local install (the free, open-source route). The official method from OpenAI's GitHub repository is pip install -U openai-whisper, which requires Python 3.8–3.11, the FFmpeg command-line tool, and PyTorch. Windows users also need to enable Developer Mode before setup works. The hardware requirement is where "free" gets complicated: the large-v3 model (1.55B parameters) needs roughly 10 GB of GPU VRAM, and running it on CPU is painfully slow. The turbo model is a middle ground — 809M parameters, about 6 GB of VRAM, and roughly 8x faster than large. If you have a decent GPU and want everything offline on your own machine, this is the route.

• Local install: `pip install -U openai-whisper`, Python 3.8–3.11, FFmpeg, PyTorch, ~10 GB VRAM for large-v3
• Turbo model: ~6 GB VRAM, ~8x faster than large — a practical compromise for smaller GPUs
• Hosted API: whisper-1 at $0.006/minute ($0.36 per audio hour), no hardware needed
• Managed models: gpt-4o-transcribe and gpt-4o-mini-transcribe (March 2025) — not open-source, not newer versions of open-source Whisper

Path 2: The hosted API. If you don't want to touch a GPU, OpenAI sells access to whisper-1 as a hosted model at $0.006 per minute — $0.36 per audio hour. For a three-hour interrogation recording, that's a little over a dollar. You upload audio, get text back, and never install anything. The trade-off is control: your audio goes to OpenAI's servers, and you're bound by their service rather than your own hardware.

Path 3: Managed models that aren't Whisper. In March 2025, OpenAI released gpt-4o-transcribe and gpt-4o-mini-transcribe, which reportedly have lower word-error rates. These are worth knowing about for one reason: they are not open-source and are not newer versions of the open-source Whisper you'd install locally. If your reason for choosing Whisper is control — running offline, keeping the pipeline on your own hardware — the managed models don't give you that.

One caveat applies to all three paths: none of them is a finished transcription workflow. Whisper turns audio into text and stops there — it doesn't label who spoke, doesn't cite timestamps you can click into, and doesn't mark output for human review. For evidence work, those gaps matter more than the install method, which is why some defense teams end up with a workspace like Evidence Copilot that handles transcription, diarization, and citation-ready review as one built-in step rather than a pipeline you assemble yourself.

Where DIY Whisper Breaks Down for Casework: Diarization, Accuracy, and Hallucinations

If you've installed Whisper and pointed it at a jail call, you've already noticed the problem: the transcript comes back as one undifferentiated wall of text. That's not a bug — it's the design. Whisper "turns audio into text, and that's the whole job. It doesn't tell you who spoke," according to AssemblyAI's diarization guide.

To get speaker labels, you have to bolt on a second model — pyannote.audio or WhisperX — plus custom alignment code to stitch the two outputs together. That pipeline has documented failure modes: it swallows short turns, collapses overlapping speech, and merges speakers on long recordings. Those are exactly the conditions you find in interrogation footage and jail calls, where exchanges are rapid, crosstalk is constant, and who said what is often the entire issue for the suppression motion. AssemblyAI's blunt summary: most teams reach for Whisper + pyannote "because it looks free and turns out to be expensive — in GPU time, in glue code, and in the accuracy you quietly lose on the hardest turns."

Accuracy is the second gap. The headline 3.6% word-error-rate comes from clean, read audio — what researchers call "a ceiling, not a forecast." On crowd-sourced speech recorded on ordinary microphones, WER jumps to 12.4%, and phone-quality audio pushes it to 28.8%. Bodycam mics, interrogation-room audio, and recorded jail lines sit much closer to those real-world conditions than to a studio benchmark.

The third gap is the one that should worry you most: hallucination. A peer-reviewed ACM FAccT 2024 study documented Whisper fabricating content during silences, with hallucination rates ranging from 1% to 80% of segments depending on conditions — worst with long silences, frequent pauses, and speech-like background noise. Picture a silent stretch in an interview room followed by a transcript that invents a statement nobody made. On cross-examination, that's a credibility problem for you, not the software.

The broader evidence points the same direction. A Northeastern study of AI transcription on 176 bodycam videos — testing Rev.ai, not Whisper — found AI identified an average of 3.3 participants per transcript versus 4.8 in human-edited versions, and the AI transcripts ran roughly 25,000 words short across the sample. As EPIC's Calli Schroeder put it: "An AI transcription or summary may be 98% accurate, but that 2% could completely change who is at fault and what occurred in an encounter."

None of this means AI transcription is useless for defense work — the same coverage notes that public defenders use it to "find the two minutes they're looking for without having to scrub through a three-hour video file." It means AI transcripts are drafts, and the review step is non-negotiable. That's the workflow Evidence Copilot is built around: diarized, timestamped transcripts where every AI output is marked pending human review, every answer cites the exact second of video it came from, and the original files stay immutable. Whether you run DIY Whisper or a managed workspace, the rule holds — the machine prepares the file, and a human decides what it means.

What 'Free' Actually Costs: The Hidden Price of a DIY Pipeline

Nobody charges you a line item for glue code, but you pay it anyway. Whisper's MIT license means no fee, yet the real cost of a DIY pipeline shows up in hardware, engineering hours, and the accuracy you lose exactly where it matters most — the hardest turns of a jail call or interrogation recording.

The hardware bill arrives first. The large-v3 model needs roughly 10 GB of GPU VRAM to run, and without one, transcription is painfully slow on CPU. A GPU is effectively required for medium and large models at any real speed, per AssemblyAI's own deployment guidance. Running 40 hours of bodycam through a consumer laptop is not a weekend project; it is a hardware procurement decision.

Then comes the code Whisper does not ship. The model turns audio into text, and that is the whole job — it does not tell you who spoke. Adding speaker labels to an interrogation video requires a three-stage pipeline: Whisper transcription, a separate diarization model like pyannote, and custom alignment code you build and maintain yourself. That pipeline has documented failure modes:

• Short interjections get swallowed — the "yeah," the "no, sir," the admissions buried in quick exchanges.
• Overlapping speech collapses, since Whisper transcribes a single stream and does not separate speakers.
• Speaker labels merge on long recordings, scrambling who said what across a two-hour interview.

The vendor math is blunt. As AssemblyAI puts it, "most teams reach for Whisper + pyannote because it looks free and turns out to be expensive — in GPU time, in glue code, and in the accuracy you quietly lose on the hardest turns." A catastrophic speaker-swap output can score 15.1% on one metric and 30.7% on another — the same transcript, two very different verdicts, and no easy way to know which one you are looking at.

There is also no audit trail, no redaction tooling, and no citation layer. When you find the incriminating minute in a DIY transcript, you still have to hand-map it back to the source video yourself. And the accuracy floor is lower than the benchmarks suggest: on crowd-sourced speech recorded with ordinary microphones, Whisper's error rate was 12.4% — roughly three and a half times worse than its headline figure — with phone-quality audio adding about seven more points of error.

The one genuine DIY advantage is control. Running locally keeps discovery on your own hardware with no third party in the loop, and for practitioners with air-gapped requirements, that matters. If that describes your practice, budget honestly for the pipeline above.

For everyone else, the honest comparison is DIY Whisper against a workspace where transcription, diarization, citations, and redaction already exist. Evidence Copilot's built-in transcription uses Deepgram Nova with speaker diarization and word-level timestamps, and every AI answer cites the exact video second or PDF page it came from — with originals kept immutable and AI output always marked pending human review. If you want to see that workflow against your own discovery, book a 20-minute walkthrough at evidencecopilot.com/book, or start on the Solo plan at evidencecopilot.com/signup.

The Ready-Made Alternative: Transcription Built Into Evidence Review

There's a third path that skips the pipeline entirely: transcription that's already built into the evidence-review workspace. For a solo defender staring down 40 hours of bodycam before a preliminary hearing, the question isn't really "how do I install Whisper" — it's "how do I get a transcript I can actually use."

That's the gap Evidence Copilot fills. It uses Deepgram Nova transcription with speaker diarization and word-level timestamps, so interrogation footage and jail calls come back already separated by speaker — no pyannote glue code, no alignment scripts to maintain. That matters because the DIY approach has documented failure modes: research on Whisper-based diarization pipelines shows they swallow short turns and collapse overlapping speech, the exact conditions you find in a three-person jail call. One vendor analysis put it bluntly: most teams reach for Whisper plus pyannote "because it looks free and turns out to be expensive."

The workflow problems Whisper leaves open are where the built-in approach earns its keep:

• One-click speaker renaming that propagates across every segment — fix "Speaker 2" to "Ofc. Ramirez" once, done.
• Every AI answer cites the exact source moment it came from, with a clickable citation that seeks the precise second of video or page of the PDF.
• All AI output is marked AI-generated and pending human review, with originals kept immutable.
• Full-text search across transcripts and documents, so you find the two minutes you need without scrubbing a three-hour file.

That last point is the real payoff. Coverage of an AI-transcription study of 176 bodycam videos noted that public defenders value AI transcription precisely because it lets them "find the two minutes they're looking for" in hours of footage — but the same study found AI identified 3.3 participants per transcript versus 4.8 in human-edited versions. The searchable record is the benefit; the human review is non-negotiable. Experts quoted in that coverage warned about automation bias, and an EPIC analyst put it plainly: a transcript that's 98% accurate can still get the 2% wrong that changes what occurred in an encounter.

On price, there's no sales wall. The Solo plan runs $39/user/month with 15 media-processing hours included; Firm is $99/user/month with 60 hours and a priority queue; the PD Office plan is $750/attorney/year with 100 hours — roughly 38% under JusticeText's published $1,200/attorney/year rate. Overage is metered and hard-capped at 3x your included hours, so a surprise discovery dump can't turn into a surprise invoice. Every plan includes the full workspace; there are no feature tiers.

So the choice comes down to what your time is worth. If you want ownership of the pipeline, the local Whisper route is documented and real. If you want the searchable, cited, review-ready transcript without building it, book a 20-minute walkthrough at evidencecopilot.com/book, or start on the Solo plan at evidencecopilot.com/signup.