The best AI transcription tools in 2026

One hour of audio costs about 36 cents to transcribe through OpenAI’s API. The same hour costs $119.40 through a human service charging $1.99 a minute. Both prices are real and both make sense for somebody, which is why so much confusion in this category comes from comparing tools built for different jobs.

This list is for anyone who needs words out of audio and doesn’t want to spend a week finding the right tool. That covers freelancers recording client calls, small teams who live on Zoom and Google Meet, YouTubers who need captions, and developers wiring transcription into an app. I’m a Singapore-based operator running a few small sites and content pipelines, so my lens is practical. What does it cost per audio hour? Where does the recording go? Does it cope when a speaker slides from English into Mandarin halfway through a sentence?

One caveat before the picks. I did not run a fresh word error rate benchmark across these seven, and I’d be wary of anyone who claims a clean ranking after a weekend of testing. Prices and limits come from each vendor’s public pages as I last read them. They move often, so check the linked page before you pay.

how I picked

Here is what I weighed, roughly in this order.

cost per audio hour at the volume you’d really use, not the headline monthly price. A $17 plan with a 1,200 minute cap is a different product from a per-minute API.
language coverage, with extra weight on code-switching. Calls in Singapore often mix English, Mandarin and Malay in one breath.
speaker labels and timestamps. A transcript that can’t tell you who said what is a wall of text.
where the audio goes, how long it stays, and whether it can be used for training. I’m not a lawyer, so read each vendor’s terms yourself.
how it fits the workflow: meeting bot, file upload, API, or an app on your own machine.

I left out Google Cloud Speech-to-Text, Azure AI Speech and Amazon Transcribe. They’re cloud platform services more than standalone tools, and if you live on one of those clouds you’ll likely use its own.

the picks

OpenAI Whisper and the OpenAI transcription API

Whisper is the open model OpenAI released in September 2022. The paper says it was trained on 680,000 hours of audio and covers roughly 100 languages, and the code and weights sit on GitHub under an MIT licence. You can run it on your own machine for nothing, or pay OpenAI per minute through the hosted API, which offers whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe with uploads capped at 25 MB a file.

Whisper has no speaker labels of its own, and it makes things up. The Associated Press reported in October 2024 on Whisper inserting text nobody said, including in medical settings. Read the output before anything with consequences depends on it. If you self-host, running a model on your own hardware covers what that really takes, and a smaller checkpoint can be enough for clean single-speaker audio, the same logic as in when a small model is the right choice.

pros:

free to self-host, with the code and weights under an MIT licence
runs on your own machine, so the audio never has to leave it
multilingual out of the box

cons:

no speaker labels, no meeting bot, no editor. you build or bolt those on
can hallucinate on silence and noisy stretches, so a human read is not optional

pricing: free to self-host, hardware costs aside. API at $0.006 a minute for whisper-1 and about the same for gpt-4o-transcribe, $0.003 a minute for gpt-4o-mini-transcribe, as of my last read.

link: OpenAI speech to text docs

Deepgram

Deepgram is a speech API for developers, and streaming is where it earns its place. If you need live captions, a voice agent or call audio transcribed as it happens, it’s built for that. Its flagship when I last read the docs was Nova-3, released in February 2025.

The API can return speaker labels, word timestamps and confidence scores. Nova-3 added multilingual code-switching, but for a short list of languages, so check yours before you assume it copes with a Mandarin and English call. Some features are billed as add-ons, so model your real bill on a week of actual audio before you commit.

pros:

real-time streaming plus pre-recorded batch, in the same API
pay as you go with no seat licences, and new accounts got $200 in free credit when I last looked
speaker labels, timestamps and per-word confidence in the response

cons:

it’s an API, so somebody has to write code. no editor, no notes page
add-on features and multilingual models push the per-minute price up

pricing: pay as you go. roughly $0.0043 a minute for pre-recorded English on Nova-3 and $0.0077 for streaming when I last checked.

link: Deepgram pricing

AssemblyAI

AssemblyAI is the other developer API here. It sells speech to text plus the things you’d otherwise build on top: speaker labels, summaries, topic and sentiment detection, and PII redaction. Redaction is the one that matters for support calls, where names and card numbers turn up. Check its language table for yours before you commit.

The base rate is the lowest hosted price here by my last read, but add-ons are billed separately, so a pipeline with speaker labels, summaries and redaction won’t cost the headline figure. Using the summary features also means your transcript passes through another model, and another set of terms. I’d start here over Deepgram for batch jobs on recorded calls, and lean to Deepgram for live audio.

pros:

lowest base price among the hosted options, about $0.15 per audio hour by my last read
PII redaction, speaker labels and summaries from the same API
official SDKs, Python and JavaScript among them

cons:

add-ons stack, so the headline price is a floor and not the bill
summaries and similar features put your transcript through another model

pricing: pay as you go. roughly $0.15 per audio hour for the base Universal model at last check, with speaker labels and other features priced on top.

link: AssemblyAI pricing

Otter.ai

Otter is the one I’d hand to someone who doesn’t want to think about any of this. You connect your calendar, its bot joins Zoom, Google Meet or Microsoft Teams calls, and a transcript with a summary and action items shows up afterwards. It also transcribes live during the meeting, which helps when you’ve lost the thread and want to scroll back.

Otter’s language list is short, mostly English with a few others when I last looked, so it’s a poor fit if your calls regularly run in Mandarin or Malay. Pro at 1,200 minutes a month is 20 hours, which a busy consultant will hit. The bot also announces itself in the meeting. That’s the right behaviour, but tell people before you send it into a client call.

pros:

easiest setup on this list: connect a calendar and you’re done
live transcript during the call, summary and action items after
a free tier that’s enough to find out if you’d use it, 300 minutes a month at last check

cons:

short language list, a weak fit for mixed English and Mandarin calls
monthly minute caps that a heavy user will hit

pricing: Basic free (300 minutes a month, 30 minutes per conversation). Pro $16.99 a month, or $8.33 a month billed annually. Business $30 a month, or $20 annually, per user. as of my last read.

link: Otter pricing

Fireflies.ai

Fireflies does the same job as Otter, a bot that records and transcribes your calls, and it competes on breadth. The vendor claims support for more than 100 languages, and it plugs into CRMs, Slack and project tools so notes land where the team already works. If your calls run in more than one language, put your own audio through the free plan before you believe that claim.

My take, and people will argue with it: the meeting-bot apps are a speech model, a calendar bot and a summary prompt in a nice wrapper. That’s worth paying for if nobody on your team will build it. It’s worth less if you’re happy with a script and a Whisper install, because seat pricing grows with headcount while API cost grows with audio. With 15 people on calls, seats are easier to justify than engineering hours.

pros:

claims 100+ languages, a far wider list than Otter’s
integrations with CRMs, Slack and other work tools
a free plan that includes transcription, with storage and AI summaries limited

cons:

per-seat pricing, so the bill follows headcount and not usage
the bot joins as a visible participant, which some clients dislike

pricing: free plan. Pro $18 per seat monthly, or $10 billed annually. Business $29 monthly, or $19 annually. as of my last read.

link: Fireflies pricing

Descript

Descript solves a different problem. It transcribes audio or video, then lets you edit the media by editing the text: delete a sentence in the transcript and it’s cut from the recording, and captions export as SRT. If you make podcasts or YouTube videos, the transcript is a stepping stone to a cleaner edit, and that’s where Descript earns its price. I wouldn’t buy it just to get a transcript file.

Its plans and allowances have been reshuffled more than once, so treat any older comparison, this one included, as a starting point and read the current tier table. There’s a free plan to try. It’s also a full editor rather than an upload box, so expect a learning curve.

pros:

edit audio and video by editing the transcript
filler word removal and captions built in
record, transcribe, edit and export in one app

cons:

overkill if all you want is a transcript file
allowances and pricing have changed repeatedly, so budgeting is fiddly

pricing: free plan with limits. Paid tiers from about $16 a month per person billed annually when I last looked, more for higher tiers.

link: Descript pricing

Rev

Rev is here because it puts a human behind the transcript. Its human service is listed at $1.99 a minute and the company advertises 99% accuracy for it. It also sells AI-only transcripts for people who want a cheap first draft. I haven’t compared those plans, so check the current tiers.

Most people can’t justify $119 an hour, and I wouldn’t spend it on routine meetings. But a 20-minute interview you’re going to quote in print, or a recording that might turn up in a dispute, costs about $40 to have checked by a person, and that’s cheap next to a misquote.

pros:

a person checks the text, which nothing else on this list offers
a stated accuracy standard, 99% advertised for the human service
one vendor for both AI drafts and human transcripts

cons:

$1.99 a minute is more than 300 times OpenAI’s $0.006 for whisper-1
turnaround isn’t instant, so no use for live captions or quick jobs

pricing: human transcription listed at $1.99 a minute at last check. AI plans priced separately.

link: Rev

comparison table

Prices are as I last read each vendor’s page.

tool	price	primary strength	primary weakness
OpenAI Whisper and API	free to self-host, API from $0.003 a minute	free, self-hostable, wide language coverage	no speaker labels, can hallucinate
Deepgram	about $0.0043 a minute pre-recorded	real-time streaming API	needs developer time
AssemblyAI	about $0.15 per audio hour base	PII redaction and speaker labels	add-ons inflate the bill
Otter.ai	free, Pro from $8.33 a month annual	easiest meeting notes	short language list, minute caps
Fireflies.ai	free, Pro from $10 a seat annual	wide language claims, integrations	per-seat cost, visible bot
Descript	from about $16 a month annual	edit media through text	overkill for plain transcripts
Rev	$1.99 a minute for human transcripts	a human checks the text	300 times the cheapest API

how to choose

Start with the job, not the tool. Text from a folder of recordings points to Whisper or AssemblyAI, meeting notes without code to Otter or Fireflies, live captions or a voice feature to Deepgram, an edit to Descript, and high stakes to Rev. Most disappointment comes from buying the right tool for the wrong job.

Then test with your own audio. Take ten minutes of your messiest real call, the one with crosstalk, product names, numbers and some code-switching, and run it through two or three candidates. Count the errors that matter, which are usually names and figures rather than filler words. I’d ignore vendor accuracy percentages, because they’re often measured on clean test sets that look nothing like a real call. ai benchmarks and why they mislead makes the general case, and speech models and where they still fall over covers the failure patterns to look for. I haven’t tested all seven on Singapore-accented English or on Mandarin and English mixed in one sentence, and the vendor language lists don’t settle it.

Next, where does the audio go. Meeting bots and hosted APIs mean a third party holds a recording of your calls, at least for a while. Read each vendor’s retention and training terms, and if a call involves clients, tell them it’s being recorded. In Singapore the Personal Data Protection Commission publishes guidance on the PDPA at pdpc.gov.sg, and it’s worth reading before you put a bot on client calls. This is not legal advice. Self-hosting Whisper is the option that keeps audio on your own machine. The Privacy Wire, our sister site, writes about privacy more broadly.

Last, do the sum at your real volume. Say you record 20 hours of calls a month. OpenAI’s whisper-1 comes to about $7.20, AssemblyAI’s base rate to about $3 before add-ons, Otter Pro to $16.99 on monthly billing (right at its 1,200 minute cap), and Rev’s human service to $2,388. Same audio, wildly different bills. More posts like this live on the blog.

verdict / top pick

My top pick is OpenAI Whisper, run yourself or through the API, for anyone comfortable with a script. It’s free to self-host, $0.006 a minute to rent, it covers roughly 100 languages, and there’s no seat licence or bot to manage. You accept two costs: you build the speaker labels and summaries yourself, and you read the output because it can invent text.

If you’d rather not touch code, Otter is the easiest start for English-language meetings and Fireflies is the one to trial if your calls span languages.

If one tool beats Whisper on your own worst ten minutes of audio, ignore my ranking and use that one.

Written by Xavier Fok

disclosure: this article may contain affiliate links. if you buy through them we may earn a commission at no extra cost to you. verdicts are independent of payouts. last reviewed by Xavier Fok on 2026-09-24.