The Two-Path Answer: API Costs Money, the Model Doesn't

The short answer is yes and no, depending entirely on which "Whisper" you mean. The distinction trips up a lot of defenders who Google "free transcription" while staring down 40 hours of bodycam footage.

The Whisper API has no free tier. OpenAI bills from the first minute at $0.006/min ($0.36/hour) for the whisper-1 model, and you need an account, an API key, and a credit card on file before you transcribe a single file. A 10-hour proof-of-concept run costs $3.60 — small money, but it's never zero, per published pricing breakdowns.

The API also carries a 25 MB file size limit per request. A 28-minute MP3 at 128 kbps runs roughly 27 MB, so a single bodycam file from an ordinary traffic stop can already exceed the cap. Anything longer — a full shift, a jail call log, an interrogation recording — requires chunking and stitching the audio yourself.

The open-source Whisper model is a different story. It's free to download and run on your own hardware, with no usage fees and no API limits. A practitioner running it locally on an RTX 3090 processed a 42-minute recording in about two minutes, with no audio leaving the machine and no cloud calls, according to a hands-on writeup. For a solo defender worried about client confidentiality, that privacy angle is real.

But "free" is doing heavy lifting in that sentence. Multiple sources converge on the same caveat: the model costs nothing, yet self-hosting requires GPU hardware, engineering time, and maintenance. The DIY stack for speaker labels — Whisper plus pyannote plus WhisperX — takes 2–4 hours of setup, Python 3.8–3.11, 8GB+ RAM, ffmpeg, and a HuggingFace account where you manually accept terms for three separate gated models. Full self-hosted costs, once DevOps and storage are included, run an estimated $861–1,500/month, with break-even only around 2,400 hours/month.

And here's the part that matters most for evidence work: Whisper produces no speaker labels. It emits a continuous transcript without indicating who spoke when. For a podcast, that's an inconvenience. For a jail call where the State claims your client made an admission, "who said what" is the entire ballgame — as diarization research for legal teams puts it, transcription answers what was said, while legal work runs on who said it. Even adding a diarization layer doesn't fully solve it: overlapping speech gets collapsed to one speaker, and short turns get swallowed by neighboring speakers on long recordings.

So the honest two-path answer looks like this:

• API route: no free tier, $0.006/min from minute one, 25 MB file limit, batch-only, no speaker labels.
• Local open-source route: genuinely free, unlimited use, full privacy — but you supply the GPU, the glue code, and the diarization QC.
• Either way: no citations into exact video moments, no redaction, no audit trail — the things case work actually requires.

That last gap is where a purpose-built workspace earns its keep. Evidence Copilot's Solo plan is $39/user/month with 15 media-processing hours included — published pricing, no sales wall — and includes diarized transcripts with word-level timestamps and one-click speaker renaming, so the "who said what" layer comes built in rather than bolted on. If you want to see how that compares against a weekend of DIY Whisper setup on your own discovery, you can book a 20-minute walkthrough at evidencecopilot.com/book or start on the Solo plan at evidencecopilot.com/signup.

What 'Free' Actually Costs: The DIY Whisper Stack's Hidden Bill

"Free" is the most expensive word in this conversation. The open-source Whisper model costs nothing to download — but a practitioner's walkthrough of a local Whisper pipeline makes clear what you actually pay: a GPU, hours of setup, and ongoing maintenance of glue code before the first transcript appears.

The honest starting point is that a fully local stack does work. One practitioner ran Whisper with speaker diarization on an NVIDIA RTX 3090 — no cloud APIs, no subscriptions — and transcribed a 42-minute recording in about two minutes. But that setup required Python 3.8–3.11, ffmpeg, and accepting terms on three separate gated HuggingFace models before anything ran. Initial setup time: 2–4 hours, and that's before you've reviewed a single bodycam file.

Then there's the hardware. Peak VRAM during diarization runs around 8 GB, and the full Docker image with all models weighs roughly 20 GB. If you're running on CPU instead of GPU, expect 1–2 hours of processing per hour of audio — which turns a 40-hour discovery dump into a multi-day batch job.

The real bill shows up when you scale past hobbyist use. According to a 2025 cost analysis, a self-hosted Whisper deployment runs $861–1,500 per month once DevOps, storage, and monitoring are included — and self-hosting only breaks even at roughly 2,400 hours of audio per month. At 200 hours monthly, the paid API is 12× cheaper. No solo defender or PD office is processing 2,400 hours a month; most are reviewing tens of hours of bodycam, jail calls, and interrogation footage per case.

And the stack still doesn't do the job. Whisper produces a continuous transcript with no indication of who spoke when — and as one diarization analysis puts it, "legal work runs on who said it." Your DIY pipeline also gives you:

• No citations linking a quote to the exact second of video or page of a PDF
• No redaction tooling for discovery turnover
• No audit trail when opposing counsel asks how you produced an exhibit
• Diarization quality-control gaps — overlapping speech collapsed to one speaker, short turns swallowed

The trade is real: dollars for hours, hours for gaps. If you'd rather spend those hours on the motion to suppress than on maintaining alignment code, a workspace like Evidence Copilot starts at $39/user/month with 15 media-processing hours included and published pricing — no sales wall, no GPU.

The Diarization Gap: Whisper Doesn't Tell You Who Said It

Run Whisper on an interrogation recording and you get accurate words with no attribution — a wall of text where the detective's questions and your client's answers blur into one voice. For a podcast, that's an inconvenience. For a motion to suppress or a report-vs-bodycam impeachment, it's fatal.

Whisper produces a continuous transcript with no speaker labels at all — it doesn't indicate who spoke or when, according to the standard diarization guide for Whisper. As one vendor in the space puts it, transcription answers what was said, but legal work runs on who said it. In report-vs-bodycam impeachment, "who said what" is the whole game: the report claims your client consented to the search, the bodycam shows the officer talking over him — if you can't attribute the turns, you can't make the point.

The free workaround is real: Whisper for transcription, pyannote.audio for diarization, WhisperX to glue them together. One practitioner documented the full local pipeline running on an RTX 3090 — no cloud, no subscriptions. But the stack needs 2–4 hours of setup, Python 3.8–3.11, 8GB+ RAM, ffmpeg, and a HuggingFace account with three separately gated models accepted by hand.

Then the failure modes show up on exactly the audio that matters most:

• Short turns get swallowed by neighboring speakers — the "yeah" or "no" that decides a consent dispute disappears into the other party's segment.
• Overlapping speech collapses to one speaker, and per analysis of the DIY pipeline, that happens constantly: overlapping speech occurs in 20–40% of turn transitions in natural conversation, with contested proceedings at the high end.
• Speakers merge on long recordings — a two-hour jail call quietly becomes a monologue attributed to one voice.

Every one of those failures lands on the contested moments — the interruptions, the talk-overs, the buried admissions — not the clean stretches. And a misattributed quote you didn't catch becomes a credibility problem on cross.

This is the gap that pushes most defenders past the DIY stack. Evidence Copilot handles the same audio with diarized, word-level-timestamped transcription, one-click speaker renaming that propagates across every segment, and citations that point to the exact second of video. Solo runs $39/user/mo with 15 media-processing hours included, and overage is hard-capped at 3× included hours, so a heavy discovery dump can't silently overbill you. The free Whisper route gives you words; it doesn't give you a transcript you can put in front of a judge.

Where the Free Stack Breaks for Case Work — and What a Workspace Adds

If you're a solo defender or investigator, local Whisper genuinely works — one practitioner ran a fully free pipeline on an RTX 3090 with "no cloud APIs, no subscriptions," transcribing a 42-minute recording in about two minutes. For personal, privacy-driven use, that's a legitimate win.

Case work is a different animal. When the "evidence" is 40 hours of bodycam, interrogation footage, and jail calls that you have to turn over, cite, and possibly redact, the free stack breaks in four specific ways:

• No citations into exact moments. Whisper hands you a wall of text. When you need to pull the exact second an officer says "I didn't read him Miranda" for a suppression hearing, you're scrubbing video by hand.
• No redaction. Discovery turnover often means blurring bystander faces or muting victim statements. Whisper produces none of that — you're stitching together separate tools with no guarantee the original file stays untouched.
• No audit trail. If opposing counsel or the court asks how a transcript was produced or altered, a local pipeline gives you nothing to point to.
• Chunking around the 25 MB API limit. If you use the hosted API instead, a 28-minute MP3 at 128 kbps already runs ~27 MB, so most real-world files need to be split and stitched — and the API has no free tier at $0.006/min.

Even the diarization workaround has quality gaps. The DIY route (Whisper + pyannote + WhisperX) takes 2–4 hours of setup, a gated HuggingFace account, and self-maintained alignment code — and as one technical analysis puts it, the pipeline "looks free and turns out to be expensive," with short turns swallowed by neighboring speakers and overlapping speech collapsed to one speaker. When "who said what" decides your cross, that's not a rounding error.

This is where a purpose-built workspace differs from a transcription tool. Evidence Copilot starts at $39/user/month on the Solo plan with 15 media-processing hours included — every AI answer carries a clickable citation into the exact second of video or page of a PDF, redactions are frame-accurate with originals kept immutable, and every action is audited. Overage is metered and hard-capped at 3× included hours, so a surprise discovery dump can't spiral your bill.

The honest framing: free Whisper costs you time and QC labor instead of dollars. If your caseload is a handful of short recordings, that trade may be fine. If it's weekly intake of bodycam and jail calls, the math — and the risk — shifts. You can see the full published pricing at evidencecopilot.com/pricing or book a 20-minute walkthrough at evidencecopilot.com/book.

Choosing Your Route: A Practical Decision Guide

Here's the honest framing: neither "free" path is wrong — they just solve different problems. The trick is matching the route to the stakes of the work in front of you.

Route 1: A single, low-stakes file. If you need one 911 call transcribed before a preliminary hearing and nobody will ever contest who said what, local Whisper is fine. The model is open-source and runs on your own hardware with no usage fees or rate limits — one practitioner ran a 42-minute recording through WhisperX on an RTX 3090 in about two minutes, entirely offline. Budget 2–4 hours of initial setup, a GPU (CPU-only runs at 1–2× audio length, so one hour of audio takes one to two hours), and a free HuggingFace account to unlock the gated pyannote diarization models. That's a reasonable trade for a one-off.

Route 2: Ongoing discovery intake. Once you're processing a steady stream of bodycam, jail calls, and interrogation footage, the DIY stack starts costing more than it saves. Three gaps matter:

• No diarization QC. Whisper produces a continuous transcript with no speaker labels, and DIY pipelines quietly fail on short turns, overlapping speech, and speaker merging on long recordings — exactly the conditions in contested interviews.
• No citations. You can't click a transcript line and land on the exact second of video, which makes report-vs-bodycam comparison and cross-exam prep slow and error-prone.
• No redaction or audit trail. Discovery turnover with protected faces, minors, or PII means frame-accurate redaction — and a defensible record of what you changed. A raw Whisper output gives you neither.

There's also the API trap if you go cloud instead of local: the Whisper API has no free tier, bills from the first minute at $0.006/min, and caps files at 25 MB — a 28-minute MP3 at 128 kbps is already ~27 MB, so long bodycam files need chunking and stitching before they'll even upload. And self-hosting at scale isn't free either; cost analysis puts all-in self-hosted Whisper at $861–1,500/month once DevOps and storage are counted, with break-even only around 2,400 hours/month.

That's the line where a workspace earns its keep. Evidence Copilot, for example, includes 15 media-processing hours on the $39/user/mo Solo plan — published pricing, no feature tiers, overage hard-capped at 3× so a heavy intake month can't surprise-bill you. Diarization, word-level timestamps, and one-click speaker renaming are built in, and every AI answer cites the exact source moment, always marked pending human review.

If you want to see whether the workspace route fits your intake volume, book a 20-minute walkthrough at evidencecopilot.com/book — no pitch deck, just your actual files. Or skip the call and start on the Solo plan at evidencecopilot.com/signup.