Podcast transcription

A transcript is the cheapest thing you can add to an episode and the one that pays for itself the most times. It makes the show accessible, it gives search engines something to read, and it turns an hour of audio into a document you can skim, quote, and cut from. This page covers what the options really cost, what accuracy means once you are the one fixing the file, and why the timestamps matter more than most people expect.

Why episodes get transcribed

Four distinct reasons, and they want slightly different things from the transcript, which is worth knowing before you pick a tool.

• Accessibility. Deaf and hard-of-hearing listeners cannot use your show without one. Neither can anyone in a quiet office, on a bad connection, or reading in a second language. This is the reason that needs no business case, and it is also the one that requires the transcript be genuinely correct rather than roughly right.
• Search. Audio is opaque to a crawler. A published transcript is a few thousand words of on-topic text per episode, in your own words, covering phrases nobody would think to put in a title. Episodes that rank almost always rank on a sentence somebody said in passing.
• Repurposing. Finding a good moment by scrubbing a waveform is the slowest possible way to work. With text you skim, mark, and jump. Every clip, quote card, blog post, and newsletter you get out of an episode is really a transcript operation.
• Captions. Most short-form video is watched muted. An uncaptioned clip is a silent clip, and captions are a transcript with timing attached.

The real options, and what they cost

Prices checked August 2026. The gap between the cheapest and most expensive route is roughly a hundred to one, so it is worth understanding what the money buys.

Human transcription services

A person listens and types, usually with a machine draft as a starting point. Expect somewhere around $1.00 to $2.00 per audio minute at the well-known services, which puts a weekly hour-long show at roughly $60 to $120 an episode, or $250 to $500 a month. Budget services that use non-native transcribers or longer turnaround run lower, often near $0.80 per minute.

What you get for it is the top of the accuracy range on exactly the material machines find hardest: overlapping speakers, thick accents, unfamiliar proper nouns, technical vocabulary. What you do not get is speed or word-level timing. Human transcripts usually arrive with timestamps every paragraph or every thirty seconds, which is fine for reading and useless for captions.

Worth it when the transcript is the deliverable: legal or medical material, research interviews, anything being published as a document in its own right. Hard to justify at weekly cadence on a conversational show.

Do it yourself with Whisper

OpenAI’s Whisper models are open source and free to run on your own machine. A modern laptop transcribes an hour of audio in something between a few minutes and half an hour depending on the model size and whether you have a GPU. Faster reimplementations (whisper.cpp, faster-whisper) narrow that considerably. There are also hosted APIs at roughly $0.006 per audio minute, which is about 36 cents an hour, effectively free compared to everything else on this page.

The honest catch is that raw Whisper gives you a transcript and not much else. Speaker labels are a separate problem: you need a diarization step to work out who spoke when, and then you have to map the anonymous speaker numbers to actual names yourself. Word timings from the base model are approximate; getting reliable per-word timing takes a forced-alignment pass on top. None of that is exotic, and all of it is an afternoon of setup plus a script you now maintain forever.

Genuinely the right answer if you enjoy this, do a few episodes a month, and want to pay nothing. Genuinely the wrong answer if the pipeline breaking on a Sunday night would stop you publishing.

Automated transcription tools

The subscription products (Descript, Otter, Rev’s automated tier, and the transcription built into podcast hosts and clipping tools) sit between $10 and $30 a month for a few hundred minutes, or a per-minute rate around $0.10 to $0.25. They run the same class of model you could run yourself and add the parts that are annoying to build: speaker labels, a correction interface, exports in the formats other tools want, and occasionally editing the audio by editing the text.

The thing to check before you buy is not the accuracy claim. It is whether you can export word-level timing, and whether the speaker labels can be renamed once and applied throughout. Those two details determine whether the transcript is useful downstream or just a document you read once.

What accuracy actually means

Every vendor quotes a percentage and none of them mean quite the same thing. The underlying measure is word error rate: insertions plus deletions plus substitutions, divided by the number of words spoken. So “95% accurate” is a 5% word error rate.

Put a number on that. A conversational hour runs about 9,000 words. At 5% error, roughly 450 words are wrong. At 2%, about 180. At 10%, which is a realistic figure for a noisy remote recording with two people talking over each other, about 900 words, or fifteen a minute.

Here is what the percentage hides: errors are not evenly distributed and they are not equally expensive. Four hundred and fifty scattered mistakes on filler words cost you nothing, because nobody reads “um” carefully. Forty mistakes concentrated on your guest’s name, your product name, and the three technical terms the episode is about will make the transcript look careless in exactly the places a reader is paying attention. A 3% error rate that mangles every proper noun reads worse than an 8% rate that does not.

Which means the useful question is never “how accurate is it” but “what does it get wrong, and how long does fixing that take.” For most conversational podcasts the honest answer with a current model is ten to twenty minutes of cleanup per episode, concentrated almost entirely on names.

Where automated transcription still struggles

Models have improved enormously and these five have not been solved. Knowing them tells you where to look during cleanup, which is most of the speed gain.

• Crosstalk. Two people talking at once is the hardest case in the field. The model tends to transcribe the louder voice and quietly drop the other, so the failure is invisible: the sentence reads fine, it is just missing half a conversation. This is also where speaker labels go wrong, because the boundary between speakers is genuinely ambiguous in the audio. Recording each participant to a separate track removes the problem almost entirely, which is the single highest-leverage thing you can do for transcript quality.
• Names. Proper nouns are unbounded. A model that has never seen your guest’s surname will produce the nearest plausible English words, confidently and consistently. Expect every uncommon name in the episode to be wrong every time it appears, which is at least predictable: fix it once with find and replace.
• Jargon and acronyms. Same mechanism. Domain vocabulary, product names, spelled-out acronyms, and anything invented in the last year land wide. Many tools accept a custom vocabulary or a prompt listing the terms in advance, and it helps a lot. If yours does, paste in your guest’s name, your show name, and the ten words the episode is about before you run it.
• Accents and non-standard speech. Accuracy varies by accent because training data does. Recognition of heavily accented English, regional dialects, and speech with a stammer or dysarthria is meaningfully worse than for the accents that dominate the corpus. If your show features speakers a model handles poorly, budget more cleanup time or pay a human, and do not assume the vendor’s headline number applies to you.
• Numbers, homophones, and formatting. Dates, currency, units, and spelled-out versus numeric forms are inconsistent. Homophones that need context (their, there, they’re; principal, principle) go wrong at a low but steady rate. These are cosmetic in a caption and embarrassing in a published transcript.

What word-level timing unlocks

Almost every transcription product produces a plain transcript. Fewer produce a timestamp for every individual word, and that difference decides what you can build on top.

With paragraph-level timestamps you know that a passage happened somewhere in a thirty second window. With word-level timestamps you know that the word “actually” started at 00:14:22.180 and ended 240 milliseconds later. Three things follow from that.

• Captions timed to the word. The karaoke style where each word highlights as it is spoken is only possible if you know when each word is spoken. Block captions that swap every few seconds can be built from coarse timing; word-timed captions cannot, and they hold attention measurably better on muted vertical video.
• Clips that start and end on a breath. Cutting to a paragraph boundary means starting a clip a beat early or a beat late, and a clip that opens on half of the previous sentence loses viewers in the first second. Word timings let a cut land exactly on the start of the first word and exactly after the last, including the trailing silence you want and none of the throat-clearing you do not.
• Search that goes somewhere. A transcript you can search is useful. A transcript where clicking a word seeks the audio to that word is a different tool. The same data drives chapter markers, interactive transcripts on an episode page, and jumping straight to the moment a topic comes up.

So when you are comparing tools, the question that separates them is not accuracy. It is whether the export contains a start and end time per word, and whether each word carries a speaker. Ask for a sample file before you commit.

Formats, briefly

• TXT. Plain reading copy. No timing. Fine for a blog post.
• SRT and VTT. Caption files, timed in cue blocks of a line or two. Universal, accepted by YouTube and every video editor, and cue-level rather than word-level unless the file is written that way deliberately.
• JSON. The full structure: every word with a start time, an end time, a confidence score, and a speaker. This is the one that matters if anything downstream is going to cut, caption, or index the audio. Everything else can be generated from it.

A practical rule: keep the JSON, generate the rest. Going from word-level JSON to an SRT takes seconds. Going the other way loses information permanently.

Cleaning up a transcript in ten minutes

This works regardless of which tool produced the file, and it is much faster than reading the whole thing.

1. Write down the proper nouns before you look at the transcript: guest name, company names, products, place names, any jargon. Usually eight to fifteen items.
2. Find and replace each one. The model is consistently wrong in the same way, so one replacement usually fixes every occurrence. This is the majority of the total error you will ever notice.
3. Read the first ninety seconds properly. Intros carry the show name, the guest introduction, and the episode premise, they are the most-read part of any transcript, and they are where a mistake does the most damage.
4. Fix the speaker labels once at the top, then spot-check the two or three places where people talked over each other. Those are the only places labels usually break.
5. Stop. Leave the scattered filler words and the occasional homophone. Nobody has ever stopped reading over a misplaced “their.”

What PodPilot360 does

Transcription is the first stage of our pipeline rather than a feature bolted onto it, because everything after it reads from the transcript. Upload one audio file and you get word-level transcription with speaker names, which is then what the rest of the run uses: the clip picker reads the text to find standalone moments, the vertical shorts get word-timed burned-in captions from the same timings, the cuts land on word boundaries, and the chapters, titles, descriptions, and drafted blog post all come out of the same pass.

You are not paying for transcription separately and you are not exporting a file into another tool, which is the part that usually goes wrong. The honest caveat is the one this whole page has been making: it is a machine transcript, so names and jargon still want the ten-minute pass described above, and if your episode is a four-way conversation recorded on one microphone in a live room, no tool on this page will save it.

Related: turning one episode into a week of content, how to start a podcast, PodPilot360 vs Opus Clip, and the clipping tools compared. Tools: episode title generator and RSS feed validator.