An Hour of Audio: Turn a Recording Into a Transcript a Client Can Use

You have a 60-minute recording: a client interview, a podcast episode, a board meeting. What the client wants back is not "a transcript". It is 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 a video to cut, and nothing in it that makes them stop and rewind. This page gets you from the audio file to that document by two routes, one entirely in a browser and one that is a single command, then covers the ten minutes of editing that separate machine output from something you can send, and only then the money.

What you hand over

At the end you should have two files. A plain-text transcript with speaker names, paragraphs broken where the conversation turns, and the proper nouns fixed. And an SRT with the same words against timestamps, for whoever edits the video or needs to find minute 43. The second file is free once you have the first, and it is the one clients come back asking for.

If they also want a summary or show notes, that is a second, cheaper job: a text model reads the transcript and drafts them, and the drafting page prices that step by the token.

Step one: a file the tool will take

OpenAI's transcription endpoint accepts mp3, mp4, mpeg, mpga, m4a, wav and webm, and a file can be up to 25 MB. ElevenLabs accepts audio and video up to 3 GB and 10 hours. That difference decides the first step. Bitrate in kilobits per second, times seconds, divided by eight, is the size in kilobytes: sixty minutes of MP3 at 128 kbps comes to roughly 58 MB, more than twice OpenAI's ceiling. Either re-encode the file at a lower bitrate, or cut it into pieces under 25 MB. OpenAI's own guide adds the rule that matters: do not split in the middle of a sentence, because the model loses the context on both sides of the cut. For ElevenLabs, a one-hour MP3 goes up as it is.

Path A: in the browser with ElevenLabs Speech to Text

In the ElevenLabs dashboard, open the Speech to Text page and click Transcribe files. The upload window asks three things. Set the primary language if you know it, or leave it on Detect and let the model find every language in the audio. Switch on Tag audio events if you want laughter and applause marked in the text; for a meeting transcript, leave it off. Then add keyterms: the names of the people in the room, the company, the product, the acronyms, up to 1,000 words or phrases. The model uses them to decide how to spell what it hears, and the product guide notes that keyterm prompting raises the cost by 20%. Click Upload files.

Scribe v2 is the default model on the website. A file longer than 8 minutes is cut into four pieces and transcribed in parallel, so an hour does not take an hour. When it is done, click the file name in the centre pane. Every word is clickable and plays the audio from that point, speakers are separated automatically (up to 32 of them), and each word carries its own timestamp. English, French, German, Spanish, Portuguese and about thirty other languages sit in the band the docs call 5% word error rate or below; Vietnamese is in that band too.

Path B: one request to OpenAI's transcription endpoint

The API route is one request to /v1/audio/transcriptions. OpenAI's guide says to start with gpt-transcribe, its recommended model for recorded speech in its original language, and to reach for a specialised model only when you need speaker labels, word timestamps, subtitle formats or translation into English. With the official command-line client the whole job is:

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

Three extra fields do the work the keyterms did on the other route. prompt takes a sentence about the recording ("A quarterly review between the founder and the head of sales"). keywords takes the literal terms you expect to hear. languages takes the language codes spoken. The guide is careful about the middle one: keywords are hints, not required output, so include only terms that are actually in the audio, or unspoken words can appear.

For speaker labels, switch the model to gpt-4o-transcribe-diarize, ask for the diarized_json response format and, for anything over 30 seconds, set chunking_strategy to auto. You can attach up to four short reference clips of 2 to 10 seconds with the speakers' names, and the segments come back labelled with those names instead of Speaker 1 and Speaker 2. For an SRT or VTT with timestamps, use whisper-1 with timestamp_granularities; the guide says that parameter is supported only on whisper-1, and that whisper-1 takes a prompt of at most 224 tokens. If the recording is in another language and the client wants English, there is a separate /v1/audio/translations endpoint, whisper-1 only, English only.

The ten minutes that make it deliverable

Machine output is a draft. The edits that turn it into a deliverable are always the same four.

Rename the speakers. In the ElevenLabs Transcript Editor, click edit next to Speakers and type the real names; every segment updates. If a segment landed on the wrong person, click the orb beside the speaker name and move it, or use Move Segments To to move all of one speaker's lines to another.

Fix the proper nouns. Search the transcript for each name and product on your list and click the word to hear it. If you did the keyterms step properly there will be few. On the API route, OpenAI's guide suggests a second pass through a text model with your spelling list, and then a check of the corrections against the audio so the model does not change what the speaker actually said.

Fix the breaks. Press Enter inside a segment to split it where the topic turns; use merge segments to join two adjacent segments from the same speaker. After edits, click align words so the word timestamps are recomputed and the SRT stays in sync.

Decide on verbatim. Scribe v2 has a no_verbatim mode that drops filler words and false starts; a readable summary transcript wants it on, a legal or research transcript wants it off. Ask the client once and write the answer down.

Export, then hand over

The export button in the top right of the Transcript Editor offers Plain text, JSON, HTML, SRT and VTT. Take Plain text and SRT. On the API route, gpt-transcribe returns the text as JSON, and the SRT comes from the whisper-1 call. Download everything the same day, for a reason that comes up under the terms below.

What the hour costs

The chart above is the whole price question in one picture. By API, an hour of audio 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 you type.

The browser route is priced differently, and it is worth doing the arithmetic once. Inside the ElevenLabs app every product draws on one pool of credits, and Speech to Text costs 330 credits a minute, so an hour is 19,800 credits. The Free plan's 10,000 credits a month cover about 30 minutes; a one-hour file does not fit. Starter, at $6 for 30,000 credits, covers about 90 minutes, so a single hour uses two-thirds of the month, roughly $3.96 of the $6. Creator is $22 for 121,000 credits, about six hours; Pro is $99 for 600,000, about thirty. The same hour through the ElevenLabs API is $0.22 in dollars, outside the credit pool. If you transcribe more than 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.

One more published number belongs on the same chart. ElevenLabs' Productions team will have a native speaker review a transcript for you, from $2.00 per minute of audio, which is $120 for the hour. That is what a company charges a client for human work on this file, and it is the price you are competing with when you quote for the ten minutes of editing above.

What a person is paid for the same hour

If you are the one typing rather than the one uploading, the platforms publish their side too. TranscribeMe pays beginning General Transcribers $15 USD per audio hour, per audio hour rather than per hour worked, so a 2-minute file pays $0.50. Rev pays $0.40 to $1.10 per audio or video minute for transcriptionists, $24 to $66 for the hour, and $0.54 to $1.10 for captioners. Both pay through PayPal and nothing else; TranscribeMe says without a PayPal account it will not work with you, needs a $10 balance and a withdrawal before 9:00 AM PST on a US Thursday, and pays once a week. Rev pays weekly with no stated minimum, but its freelancer page says applications sit on a waitlist. Neither company publishes how many working hours an audio hour takes, so neither publishes an hourly wage, and this page does not invent one.

Who owns the transcript, and what is deleted

For the API route, OpenAI's services agreement assigns to the customer all of its right, title and interest in the Output, and the customer keeps ownership of the Input. Recording in, transcript out, both yours. The same agreement says that when it terminates OpenAI deletes all Customer Content within thirty days and hands nothing back, which is why the export step says the same day.

For the browser route, the line that matters is the licence. ElevenLabs' free plan does not include a commercial license and cannot be used for any commercial purpose, and free-plan output must credit elevenlabs.io. A transcript a client pays you for is commercial use, so the Free plan is for your own recordings; every paid plan includes the licence, and the API is billed in dollars outside the plans altogether. The voiceover page reads the same licence line for generated speech, where it bites harder.

What to do today

Take one recording you already have. Write the list of names and terms first. Upload it to ElevenLabs Speech to Text with those keyterms, or send it to gpt-transcribe with them as keywords, and time yourself from upload to exported SRT. That number, plus the machine cost from the chart, is your cost for the job. Quote against $120, which is what the tool's own maker charges for a human to do the same hour.