A Company Charges $120 to Transcribe an Hour of Audio. The Machine Does It for $0.22

Why this pays

The same hour of audio has four published prices, and they are two orders of magnitude apart.

Read it from the bottom up. The machine now types the hour for 22 cents. The person who used to type it is paid $15 to $66. The company that sells the finished transcript charges $120. The service that fits in that gap is ten minutes long: take the machine's draft, put the right names on the right lines, export the two files the client actually opens, and quote per audio minute.

That is the whole play. The rest of this page is how to do it so the client never has to rewind.

The play, in three lines

1. Upload the recording to ElevenLabs Speech to Text in the browser, or send it to OpenAI's transcription endpoint with one command.
2. Spend ten minutes renaming the speakers and fixing the proper nouns, then export a text file and an SRT.
3. Quote per audio minute, working up from the machine cost, not down from a plan price.

The path below walks every click with screenshots from a real account. What follows is the reasoning behind each step and the numbers you quote from.

Where the money comes from

Four buyers for the same ten minutes, in the order they pay:

1. People who own recordings. Podcasters, interviewers, researchers, anyone who records meetings. They pay per audio minute for the document, and the ceiling is the $2.00 a minute a company charges for human review. List it as a Fiverr Gig or send Upwork proposals; the freelance page covers the fees and the first payout.
2. Video editors who need captions. Same file, exported as SRT. If they also want other languages, the subtitles page adds translation by the token.
3. The same clients, a second time. Show notes, a summary, a blog post from the transcript: a text model drafts them and the drafting page prices that job.
4. The platforms, as a floor. TranscribeMe pays $15 per audio hour and Rev $0.40 to $1.10 a minute for typing. If a client offers less than that for the finished document, the client is the wrong buyer.

The pain you are selling the cure for

The client does not want "a transcript". They want a document with the right people's names on the right lines, the product names spelled the way the company spells them, timestamps if there is video to cut, and nothing that makes them stop and rewind. Machine output on its own fails all four. Human typing passes them at $2.00 a minute, and slowly.

So the thing you sell is not transcription. It is the ten minutes that turn machine output into that document, delivered the same day. Price it as that.

Before you upload: the file

• OpenAI's endpoint takes mp3, mp4, mpeg, mpga, m4a, wav and webm, up to 25 MB per file. Sixty minutes of MP3 at 128 kbps is roughly 58 MB, so either re-encode at a lower bitrate or split at a pause, never mid-sentence.
• ElevenLabs' upload window says audio and video files up to 1000MB; the API behind it takes up to 3 GB and 10 hours. A one-hour MP3 goes up as it is.
• Write the list of names, places and product terms before you upload. It goes in as keyterms on ElevenLabs, or as prompt and keywords on OpenAI, and it is the difference between ten minutes of editing and an hour.

Path A: ElevenLabs in the browser

Everything here was walked through in a real ElevenLabs account in 2026; the screenshots in the path below are from that session.

• Speech to Text is in the left menu. Click the black Transcribe files button, top right.
• The window has four tabs (Upload, Record, YouTube, URL) and six controls: Primary language, Tag audio events, Include subtitles, No verbatim, Assign speakers from library, Keyterms.
• Tag audio events is on by default and marks laughter and applause; switch it off for a meeting transcript. No verbatim drops filler words; decide that with the client, because it has to be set before the upload.
• Attach the file and a line appears beside Upload files with the credit cost of this exact file. Read it before you click; it is the only place the app quotes the price in advance.
• A half-minute file came back in under ten seconds. The docs say a file over 8 minutes is cut into four pieces and transcribed in parallel, so an hour does not take an hour.
• Open the transcript from the list: every word plays the audio from that point, speakers are separated (up to 32, per the docs), and each segment carries its times.

Path B: one OpenAI call

openai audio:transcriptions create --model gpt-transcribe --file meeting.mp3 --raw-output --transform text

• gpt-transcribe is the model OpenAI's guide recommends for recorded speech. Pass prompt (a sentence about the recording), keywords (the literal terms you expect to hear) and languages (the codes spoken). Keywords are hints, not required output, so include only terms that are really in the audio.
• For speaker labels, switch to gpt-4o-transcribe-diarize with the diarized_json response format and, for anything over 30 seconds, chunking_strategy set to auto. Up to four reference clips of 2 to 10 seconds let you name the speakers.
• For an SRT or VTT, use whisper-1 with timestamp_granularities; that parameter is whisper-1 only, and whisper-1 takes a prompt of at most 224 tokens. Translation into English is a separate endpoint, whisper-1 only.

The ten minutes that make it deliverable

Four edits, always the same four:

• Rename the speakers. They arrive as Speaker 0 and Speaker 1. Click the label beside any segment, type the real name, press Enter; every segment of that speaker takes the name. The top bar shows Last saved; there is nothing to save by hand.
• Fix the proper nouns. Click each name on your list to hear the audio from that word. With keyterms done properly there will be few. On the API route, run a second pass through a text model with your spelling list, then check its corrections against the audio.
• Fix the breaks. Enter splits a segment; merge segments joins two from the same speaker. The Segment properties panel shows start, end and duration, with Delete and Align; run Align after edits so the SRT stays in sync. Run Spell Check is one click in the same panel.
• Settle verbatim once. Readable summary transcript: No verbatim on. Legal or research transcript: off. Ask the client once and write it down.

Export, then hand over

• Export lists Text, PDF, DOCX, JSON, HTML, SRT and VTT.
• SRT and VTT have a step in front: the first time you pick one, Generate subtitles opens with Max characters per line 42, Max lines per cue 2, Max seconds per cue 7. Those are the usual broadcast conventions; change them only if the client's style guide says so. Click Generate, and a Subtitles tab shows every line's character count.
• Deliver Text and SRT; DOCX if the client edits in Word. On the API route, gpt-transcribe returns the text and whisper-1 returns the SRT.
• Download everything the same day, for the reason under the terms below.

What the hour costs

The chart above puts all nine published rates on one linear scale, so the length of each bar is the price: the company's $120 fills the width, the platforms' $15 to $66 take a fifth to a half of it, and the five machine rates are hairlines. By API, an hour is $0.18 on gpt-4o-mini-transcribe, $0.22 on Scribe v2, $0.27 on gpt-transcribe, $0.36 on Whisper and $0.39 on Scribe v2 Realtime; gpt-4o-transcribe-diarize, the speaker-label model, is $0.006 a minute like Whisper. The AI job cost calculator runs these for any length.

The browser route is priced in credits, and ElevenLabs publishes that price two ways:

• The pricing page says Speech to Text costs 330 credits a minute, so an hour is 19,800 credits. The Free plan's 10,000 credits a month would cover about 30 minutes; Starter, at $6 for 30,000 credits, about 90 minutes; Creator, $22 for 121,000, about six hours; Pro, $99 for 600,000, about thirty.
• The upload window quoted 439 credits for our 31-second file, about 14 credits a second. At that rate an hour is about 50,000 credits: more than a Starter month, nearly half of Creator.

We uploaded 31 seconds, not an hour, so we have not seen which figure bills at length. The conflict is logged above; until it is resolved, budget on the higher number, and read the estimate line before every upload. The same hour through the ElevenLabs API is $0.22 in dollars, outside the credit pool. Past a couple of hours a month, the API is the cheaper door into the same model and the app is the editor you use afterwards.

Against all of that stands the human price: ElevenLabs' Productions team will have a native speaker review a transcript from $2.00 per minute of audio, $120 for the hour. That is what a company charges a client for human work on this file, and it is the ceiling you quote under.

If you would rather be paid by a platform

The same ten minutes can be sold to a transcription platform instead of a client, at the platform's rate:

• TranscribeMe pays beginning General Transcribers $15 USD per audio hour, so a 2-minute file pays $0.50. PayPal only; at least $10 in the balance, withdrawn before 9:00 AM PST on a Thursday, paid once a week.
• Rev pays $0.40 to $1.10 per audio minute, $24 to $66 an hour, weekly via PayPal within a week of approval, after a waitlist.

Both pay for the typing, not for the client relationship. That is why the client route pays more: the $120 belongs to whoever hands the client the finished document.

Who owns the transcript, and what is deleted

• OpenAI assigns the Output to you and you keep ownership of the Input, so the recording and the transcript are both yours in the terms.
• When the agreement ends OpenAI deletes all Customer Content within thirty days and hands nothing back. Download the same day.
• The ElevenLabs Free plan carries no commercial licence and cannot be used for any commercial purpose. A transcript you are paid for is commercial use; every paid plan includes the licence, and the API is billed in dollars outside the plan.

What to do today

1. Take one recording you already have. Check the file: under 25 MB for OpenAI, under 1000MB for the ElevenLabs upload window.
2. Write the name list. Upload with keyterms. Read the credit line before you click.
3. Rename the speakers, click through the names, export Text and SRT.
4. Time yourself. That number, and $0.22, are the two you quote from.