The Discovery Dump Problem: Why Manual Transcription Doesn't Scale

It starts the same way every case does now: a discovery dump lands in your inbox — forty hours of bodycam, three interrogation videos, a stack of jail calls, and a 911 recording — all due for review before the preliminary hearing. Public defenders across 20 states report receiving hundreds of hours of digital discovery each year, and that volume is only growing.

The math on manual transcription doesn't work at that scale. One hour of interview audio takes roughly four hours to transcribe by hand, and professional services charge $1.00–$5.00 per minute with a 24–72 hour turnaround. Sending 30 hours of discovery audio to a human transcription service costs about $2,700 — per case — and you still wait days before you can search a single word.

Most defenders who try to solve this themselves reach for the free stack: a local Whisper install and NotebookLM. It works fine on a single clean recording. At case scale, it breaks down:

• No reliable speaker diarization, so a two-person interrogation comes back as an undifferentiated block of text you have to reconstruct from memory
• No citations back to the exact video moment, so every quote you pull requires re-watching footage to verify
• No redaction workflow for discovery turnover, and no audit trail showing what you touched
• No way to compare a police report against the bodycam transcript systematically — the inconsistencies that win suppression motions stay buried

There's also an accuracy trap hiding in the DIY approach. Whisper large-v3 scores an impressive 2.7% word error rate on the clean LibriSpeech benchmark but 7.44% on real-world audio — and discovery media is the opposite of clean. Background noise alone costs 10–15% accuracy, and jail calls recorded on phone mics in noisy environments can drop to 80% accuracy or less. Low-quality transcription can be worse than no transcription when you're building a cross-examination on it.

This is the gap Evidence Copilot was built for: upload the full discovery dump and get timestamped, diarized transcripts where every AI answer cites the exact second of video or page of the PDF it came from — with every output marked as pending human review. The point isn't to replace certified transcription for the official record; it's to make the hundreds of hours of review work actually survivable before your deadline.

Yes, AI Can Transcribe Interviews — But Accuracy Depends on the Audio

The short answer is yes — AI transcription of interviews is production-viable, and it's already happening at scale. Public defenders across 20 states use AI tools to process bodycam footage, interrogations, and jail calls, with offices receiving hundreds of hours of digital discovery each year, according to NLADA's training platform.

But "can it transcribe" and "can you trust the transcript" are different questions. The real-world numbers diverge sharply from vendor benchmarks. AI transcription hits 85–95% accuracy on clean audio, yet averages only about 62% on typical business audio — the noisy, multi-speaker, accented kind you actually get from jail calls and dashcam footage, per industry statistics compiled by BrassTranscripts. Human transcriptionists still benchmark around 99%.

Four factors make or break accuracy on interview audio:

• Speaker diarization — separating and labeling who said what. Without it, a two-hour interrogation becomes a block of text you reconstruct from memory; tools lacking diarization are described as effectively useless for multi-speaker interviews.
• Background noise — costs 10–15% accuracy. Recordings made with an external mic in a quiet room reach 95%+ accuracy; a phone mic in a noisy environment drops to 80% or less.
• Overlapping speech — two people talking at once is nearly impossible for AI to transcribe accurately, and even human transcribers struggle with it.
• Legal jargon and entity errors — general-purpose engines stumble on terms like "res ipsa loquitur," and aggregate word error rates hide dangerous mistakes in names, badge numbers, and legal terms. Missing an identifier is more expensive than missing a filler word.

There's also a hallucination problem worth knowing about: some models fabricate fluent text where audio is unclear, which is worse than a gap in a legal context. AssemblyAI reports 30% fewer hallucinations in its Universal-2 model compared to Whisper large-v3, and cautions against trusting any vendor's benchmark as a stand-in for your own audio.

The practical takeaway for defense work: treat AI transcripts as a first-pass review layer, not a certified record. That's how we built Evidence Copilot — every transcript carries word-level timestamps and speaker labels, and every AI output is marked AI-generated and pending human review, with the original media kept immutable. Use AI to search and review discovery media at volume; reserve certified human transcription for anything that becomes an official record.

One more thing: don't pick a tool off a leaderboard. There's no shared independent benchmark across platforms, so the gap between published accuracy and real accuracy on your specific audio — accented witnesses, overlapping speakers, wind noise on bodycam — is often larger than the gap between competing tools. Test on representative files from your own caseload before committing.

Where AI Transcripts Belong: Review Layer, Not Certified Record

Here's the uncomfortable math: AI transcription runs 50–180× faster than a human transcriber and costs a small fraction as much — roughly $0.05–0.25 per minute versus $1.00–5.00 — but it still falls well short of the ~99% accuracy benchmark that human transcription delivers. That gap is exactly why the answer to "can AI transcribe interviews?" is yes, with a condition: AI belongs in the review layer, not the certified record.

The emerging standard across legal and professional work is a hybrid model. Organizations use AI to chew through high-volume, internal work — the rough draft of a 911 call, a jail call, hours of bodycam — and reserve human transcribers for anything where an error has legal consequences. As one industry analysis puts it, human transcription remains "the standard" for court-admissible transcripts, while AI is sufficient for the 80%+ of transcription work that stays internal. Public defenders across 20 states are already running this pattern, using AI tools to process the hundreds of hours of digital discovery they receive each year.

Why not just use AI everywhere? Because aggregate accuracy figures hide the errors that actually hurt you. A transcript that reads as "90% accurate" can still mangle the three words that decide a motion:

• Names of officers, witnesses, and parties — misidentified speakers poison impeachment material
• Numbers — times, dates, quantities, and amounts where a single digit changes the story
• Legal and domain terms — general-purpose engines stumble on phrases like "res ipsa loquitur," and a tool optimized for podcasts produces output that's "unusable for anything beyond a rough draft"

There's also the hallucination problem, which is worse than an omission. When a model fabricates text — AssemblyAI defines a hallucination as five or more consecutive erroneous words — you can end up "quoting" a suspect on words they never said. That's why low-quality transcription can be worse than no transcription at all.

The practical workflow for a defense team looks like this: run AI transcription as a fast, searchable first pass over discovery media, treat every AI output as pending human review, and order certified human transcription only for the specific recordings that will become exhibits or official records. This is the design philosophy behind Evidence Copilot — every AI answer cites the exact second of video or page of PDF it came from, and every output is marked AI-generated and awaiting attorney review. The AI gets you through 40 hours of bodycam in an afternoon; the human judgment decides what becomes the record.

One last caution before you pick a tool: there is no shared independent benchmark across transcription platforms, and the gap between a vendor's published word error rate and real-world performance on your specific audio is often larger than the gap between competing tools. Test candidates on representative samples — noisy jail calls, overlapping speakers, accented speech — and score how much correction effort the output actually requires. That number, not the marketing page, tells you where AI transcripts belong in your practice.

How to Evaluate and Deploy AI Transcription on Your Own Evidence

Vendor benchmark pages will tell you a model hits 2.7% word error rate. That number comes from clean test sets like LibriSpeech — and the same Whisper large-v3 model that scores 2.7% there scores 7.44% on real-world audio benchmarks. Your jail calls and interrogation-room recordings are worse than real-world average.

So don't trust anyone's leaderboard, including ours. The consensus across transcription engineers is blunt: there is no shared independent benchmark across leading platforms, so the only test that matters is your own audio. Here's the workflow that works.

Build your own test set before you commit to any tool.

Pull 50–100 files that actually represent your intake: noisy jail calls, accented speakers, overlapping interrogation speech, bodycam recorded at arm's length from a muffled speaker. Then score the thing that costs you time — correction effort, not vendor WER. How many minutes did fixing the transcript take? Which errors were dangerous ones (a name, a number, a legal term) versus harmless filler? Aggregate word error rate hides exactly the errors that matter; a "missing a medication, claim type, or customer account identifier is expensive" problem applies equally to a codefendant's name or a Miranda timestamp.

A few things to check specifically:

• Diarization quality. Speaker separation is "the difference between a usable transcript and a block of text you have to reconstruct from memory" — a tool without it is effectively useless for multi-speaker interviews.
• Keyterm prompting. Supplying expected case names, officer names, and legal vocabulary is the single highest-return lever on hard audio. If the tool supports it, test it with your actual terms.
• Hallucination behavior. Fabricated text is worse than missing text in your context. Overly forceful verbatim prompts can cause hallucinated disfluencies — check what the tool invents when audio is unintelligible.
• Timestamps and citations. Every quote you might use in cross-exam needs to trace back to the exact second of the original recording.

One counterintuitive step: don't pre-clean your audio with client-side noise cancellation. The artifacts those tools introduce often confuse the model more than the original noise did — let the transcription platform handle noise isolation server-side.

This is the standard we built Evidence Copilot around: diarized, word-level timestamps on every transcript, and an evidence agent that answers with citations into the exact video moment or PDF page, so every AI output is marked pending human review and every quote is verifiable against the immutable original. Whatever tool you land on, hold it to that test — your files, your correction time, your eyes on the source audio before anything leaves the office.