Transcription software looks like one product and is really three. There are finished apps where you upload a file or record a call and read the result in a browser. There are services that add a human reviewer for the cases where machine accuracy falls short. And there are speech-to-text APIs, sold by the hour or minute to developers who build the transcript into something else. A buyer who mixes these up pays for a platform when an hourly rate would do, or buys an API and discovers nobody on the team can use it. This ranking judges eleven products on what each vendor states on its own site as of October 2026: how setup works, what a recorded hour costs, how data leaves, who stores it and whether it trains on it, and which kind of buyer it suits. We did not test word accuracy.
Visibility in this ranking can be paid for. Payment moves a vendor's position within the
shortlist; it never adds a vendor, and it never changes a word of the review. The largest vendors in transcription software
cannot hold places 1 to 3. How it works: placement disclosure ·
editorial process.
How we ranked these
Setup: upload a file, or write code against an API
The first question is who will operate the thing. An app such as Happy Scribe, Sonix or Amberscript needs an account and a file, and the transcript appears in a browser editor where a person fixes names and splits speakers. An API such as Speechmatics, Gladia, AssemblyAI or Deepgram needs a developer to send audio, handle results and store the text, and it has no editor at all. Trint adds a newsroom layer on top, with story-building tools that a lawyer will not touch. Descript turns the transcript into an editing surface for audio and video. We checked each vendor's pages for the languages it names, whether it records live calls, and whether setup stops at sign-up or needs a sales conversation.
The real price: hours, minutes and what the cheap tier hides
Three billing units are in use, and they do not convert cleanly. Amberscript and Sonix sell hours, either in monthly bundles or as pay-as-you-go. Happy Scribe, Otter and Notta sell minutes inside a per-seat or per-plan subscription. The APIs sell usage, at $0.15 to $0.75 per hour for machine transcription, and add speaker labels or medical vocabulary as extras. Bundled hours that expire each month punish uneven work: Amberscript says unused hours do not carry over. Human review is priced separately, at €1.75 to €1.99 per minute where published. A free plan is only useful if its limits fit a real file, and Notta's three-minute cap per conversation shows how far apart those can be.
Getting your data out: exports, storage region and training
Exports are rarely the problem; every product here produces text, and most produce captions. What differs is the audio itself. A recording of an interview or a case file is personal data, so we read each privacy policy or terms page for three statements. Where is it stored: Amberscript says only in Europe, Trint offers EU or US tenancy, Sonix and Descript state storage in the United States, and Speechmatics lists cloud regions. Is it used for training: Amberscript and Sonix say no, Happy Scribe makes it opt-in, AssemblyAI's terms allow it with an opt-out, and Deepgram's policy is silent. And can you leave: delete and export rights, which the policies describe in general terms.
Independence from the vendor: who owns the engine behind the text
Transcription differs from many software categories because several vendors own the recognition model themselves. Speechmatics, Gladia, AssemblyAI and Deepgram sell their own engines, and others buy or license one. Amberscript's policy refers to language models it relies on, and Deepgram hosts the open Whisper model next to its own. That matters when quality drops or a price rises: an app on top of a third-party engine can only pass on the change. Corporate ownership is also a factor. All eleven are independent companies, and Otter is the most widely used app but has the least favorable data policy of the apps. We credited vendors that name their engine, offer on-premises deployment, or let a customer choose where data sits.
Who it's for: a journalist, a legal team and a product team want different tools
A freelance journalist with a dozen interviews a month wants pay-as-you-go pricing and a decent editor, which points to Sonix or Happy Scribe. A university research group with ethics-board rules wants stored-in-Europe wording and a no-training clause, which Amberscript supplies. A newsroom wants a platform with captions, search and shared projects, which is Trint's strength. A podcaster wants to cut the audio while reading it, which is Descript. A product team building call analysis wants an API and cares about cost per hour and region, which is the work of Speechmatics, Gladia, AssemblyAI and Deepgram. The order below favors a small organization that buys an app, so published prices and written data terms outrank API depth.
The 11 tools, reviewed
#1 Amberscript
Machine and human transcription hosted in Europe · Netherlands · amberscript.com
EU-only storage Human option No training
Amberscript answers the question that decides most transcription purchases for a European organization: where does the audio go. The privacy policy says data is never stored outside Europe and that audio and transcripts are never used to train the language models it relies on. Prices are on the page in euros, as hours rather than abstract credits, and a human-checked tier sits beside the machine one, so a single account covers a casual interview and a court-bound deposition. A Dutch company with a published Amsterdam address, certified to ISO 27001, it earns first place on transparency.
Where it falls short
Hours do not roll over to the next month, and bought credits must be activated within 60 days, so an uneven workload wastes money. The pricing page we read showed no free tier or trial. Human transcription takes five days at the standard rate, and only about eighteen languages are offered at human accuracy against ninety-plus by machine.
Wrong for
A podcaster or journalist who wants a free plan to try before paying, or a developer who needs a speech API with streaming. Look at Happy Scribe for a free tier, or at Speechmatics for an API.
Pricing: Machine plans at €19, €29 and €49 per month for 5, 10 and 25 hours; pay-as-you-go from €10 down to €6 per hour; human transcription from €1.85 per minute. (none)
Visit Amberscript →
#2 Happy Scribe
AI transcription with optional human proofreading · Ireland · happyscribe.com
150+ languages Opt-in training Human proofreading
Happy Scribe is the broadest of the plain transcription services here. The free plan gives a ten-minute trial of the AI plus recording of meetings up to 45 minutes, and paid tiers climb from 120 to 6,000 minutes a month with the overage rate printed next to them. Language coverage of 150+ is the widest on this page, and human proofreading can be bought per file when 90 percent accuracy is not enough. The privacy policy makes training on your content opt-in at signup, with de-identification and withdrawal at any time. That mix of range, price clarity and consent terms fits a small agency or research group.
Where it falls short
Minutes are the meter, so a long interview archive burns through the Basic plan in an hour or two. The cheap annual prices need a yearly commitment. Proofreading is billed on top at more than a euro per minute, and the 65-plus proofreading languages are far fewer than the machine list. The policy defers storage details to a separate trust center.
Wrong for
A company that wants a pure developer API with per-second billing, or a legal team that needs certified court transcripts. Deepgram or AssemblyAI serve the first; Rev-style legal services the second.
Pricing: Free plan; Basic is €17 per month (€8.50 billed annually) for 120 minutes, Pro €29 (€19 annually) for 600, Business €89 (€59 annually) for 6,000; extra minutes €0.20. (free plan)
Visit Happy Scribe →
#3 Sonix
Browser-based automated transcription with an editor · United States · sonix.ai
Pay as you go No training 54+ languages
Sonix is what to pick when transcription is a task rather than a daily habit. You can pay $10 for an hour of audio and walk away, with no subscription, or take a monthly plan when the volume appears. A 30-minute trial needs no card. The policy states that Sonix does not use your content or customer data to train its machine learning or generative AI models, which is a plain sentence where many competitors write a paragraph of qualifications. The editor, speaker labels and timestamps cover the usual research and journalism work, and annual plans include a free month.
Where it falls short
Storage is outside the EEA and the UK, which the privacy policy says plainly, so a European regulated buyer has a transfer question to answer. Extra seats cost $25 a month each, which makes a five-person team expensive. Overage on subscriptions is billed at the same $10 hourly rate as pay-as-you-go.
Wrong for
A European organization that needs data kept inside the EU, or a team that transcribes hundreds of hours a month. Amberscript stores in Europe only; the API vendors lower the hourly cost.
Pricing: Pay-as-you-go at $10 per hour; subscriptions of $25, $50 and $80 per month for 5, 20 and 40 hours, with extra seats at $25 per month. (free trial)
Visit Sonix →
#4 Trint
Transcription and story-building platform for media teams · United Kingdom · trint.com
EU or US storage Newsroom tools ISO 27001
Trint was built for people who turn recordings into published stories, and it still reads that way. The homepage describes live transcription in over 50 languages, translation into more than 70, speaker recognition and a Story Builder that cuts captions and rough edits from the transcript. Customers choose between EU and US tenancy, and the platform privacy policy lists AWS regions in Ireland and Frankfurt for the EU option. The site states that algorithms are never trained on user transcripts, and the company holds ISO 27001 and Cyber Essentials. For an editorial team it is a strong fit.
Where it falls short
We could not read the plans and prices on the public site, so a buyer must log in or ask before comparing cost. The platform privacy policy is silent on model training, so the homepage promise is the only statement. The feature set leans toward media production and goes unused by a lawyer or researcher.
Wrong for
A solo researcher or small business looking for a price they can see before signing up. Sonix and Happy Scribe publish theirs, and both list a trial or free tier.
Pricing: Plans and prices are not listed on the public site; the homepage sends visitors to the app's plans page to see them. (none)
Visit Trint →
#5 Notta
AI transcription and meeting notes for individuals and small teams · Singapore · notta.ai
Free plan 58 languages Bot-free option
Notta is the cheapest route to a working transcript on this list for one person. The free plan allows 120 minutes a month and 50 uploads, Pro raises that to 1,800 minutes with five-hour recordings, and Business drops the minute cap altogether. The homepage names 58 languages, a desktop mode that records without a bot, and SOC 2, ISO 27001, GDPR and HIPAA compliance. The service terms name Notta Pte. Ltd. in Singapore, with a Japanese entity for users there. For someone who records interviews and lectures, that adds up to a low-cost starting point.
Where it falls short
The free plan caps each conversation at three minutes, which makes it a demo rather than a working tool. The terms of service grant Notta a license to process customer data and the pages we read did not say whether recordings train its models or where they are stored. Business seats are priced per user.
Wrong for
A buyer who has to document where audio is stored or whether it feeds model training. Amberscript and Sonix state both in writing.
Pricing: Free plan with 120 minutes a month; Pro $13.99 per month or $97.99 per year for 1,800 minutes; Business $27.99 per month or $199.99 per year with unlimited minutes. (free plan)
Visit Notta →
#6 Otter
Live transcription and meeting notes for US-centered teams · United States · otter.ai
Large user base Free plan Meeting focus
Otter is the default that most people have already tried. Its free plan gives 300 minutes a month with three lifetime file imports, Pro raises that to 1,200 minutes and ten imports a month, and Business removes the minute cap with four-hour meetings. It is aimed squarely at meetings, with live captions and CRM sync rather than archive work. It sits sixth because of its policy: Otter says it trains its proprietary AI on de-identified audio and transcriptions, and recordings sit with Amazon Web Services in the United States.
Where it falls short
Only three file imports for life on the free plan, and thirty-minute meetings, make it poor for interview archives. Training on de-identified recordings is described in the policy without an opt-out we could find. Storage is American. Business pricing is $30 monthly per user, high next to the per-hour services.
Wrong for
A European public body, a law firm or any group that cannot allow customer audio to feed a vendor's models. Amberscript or Happy Scribe give written assurances; the meeting-assistant ranking compares Otter with its direct rivals.
Pricing: Free plan with 300 monthly minutes; Pro $16.99 per user per month or $8.33 billed annually; Business $30 or $19.99 billed annually; Enterprise by quote. (free plan)
Visit Otter →
#7 Descript
Audio and video editor driven by the transcript · United States · descript.com
Editor Creators Free plan
Descript turns a recording into a document you edit by deleting words, and the audio or video follows. That is a different job from the others on this page, and it earns the place for creators who need a transcript only to cut a podcast or interview. The free plan gives 60 media minutes a month with watermarked 720p export; paid tiers raise the media hours and AI credits. The privacy policy says projects have no human access without consent or legal compulsion, and Google Workspace data is not used to train its models. It is strong if editing is the real work.
Where it falls short
As a plain transcript source it is overbuilt and credit-based, with AI credits counted separately from media hours. The policy allows analysis of projects to improve the service, with an opt-out, and voice data from AI Speakers can enter research datasets. Data is processed and stored in the United States.
Wrong for
A researcher, paralegal or analyst who needs transcripts in bulk and has no video to cut. Sonix or Amberscript charge per hour and skip the editor.
Pricing: Free plan with 60 media minutes a month; paid tiers (Hobbyist, Creator, Business) with 10, 30 and 40 media hours a month, starting at $16 per month; Enterprise by quote. (free plan)
Visit Descript →
#8 Speechmatics
Speech-to-text API with cloud and on-premises deployment · United Kingdom · speechmatics.com
API On-premises Regional cloud
Speechmatics sells an engine rather than an app, and it is the one here that can run inside your own network. The pricing page lists cloud regions in the EU, US and Australia, with containers and on-device processing for Enterprise customers. Rates are quoted per hour, from $0.24 for pre-recorded audio on its multilingual model, and new accounts get $100 in credit without a card. The privacy policy, filed under Cantab Research Limited at an address in Cambridge, calls it a processor of customer data acting on the customer's instructions. A developer with a data-residency requirement will find the clearest answers here.
Where it falls short
No finished editor or meeting bot, so a non-technical team has nothing to open. The Enterprise-only deployments need a sales conversation. The policy separates customer data from partner data used for research and development, so the contract needs reading to know which applies.
Wrong for
A team with no developer, or anyone who wants to upload a file and get a clean transcript the same afternoon. Happy Scribe or Sonix cover that.
Pricing: Usage-based per hour: Melia 1 at $0.24 pre-recorded and $0.40 real-time, Enhanced at $0.75 and $0.80, Standard at $0.45; $100 free credit at signup. (free trial)
Visit Speechmatics →
#9 Gladia
Speech-to-text API from a French company · France · gladia.io
API French vendor Training opt-out
Gladia is a developer product from Roubaix, in the north of France, and it states its data terms by plan. Starter costs $0.61 per hour for pre-recorded audio and $0.75 for real-time, with a €50 credit that does not reset. Growth, committed in advance, lowers async to as little as $0.20. Automatic model-training opt-out arrives with Growth, and Enterprise adds default opt-out and zero data retention. Its privacy notice lists a French SAS and a Delaware subsidiary and says training datasets are public ones without customer recordings. For a European developer buying an API, that is a sound set of answers.
Where it falls short
The pricing page we read did not say where data is hosted or which regions are offered. Training opt-out is automatic only from the Growth plan upward, so Starter customers must act themselves. It offers no interface for end users.
Wrong for
An office manager who wants to drag in a recording and read the result. Amberscript or Sonix offer an editor; Speechmatics lists regions explicitly.
Pricing: Starter pay-as-you-go at $0.61 per hour async and $0.75 real-time with €50 free credit; Growth from $0.20 async; Enterprise by quote. (free trial)
Visit Gladia →
#10 AssemblyAI
Speech-to-text API for building voice products · United States · assemblyai.com
API Lowest rates EU endpoint
AssemblyAI prices speech at a level no app on this page matches: $0.15 per hour for Universal-2 pre-recorded audio, add-ons priced separately per hour. New accounts receive $50 in credit without a card. The pricing page states that the EU region costs the same as the US one, with separate API hostnames, which makes data residency a configuration setting and not a sales call. The company is American, with a New York address in its terms. It ranks in the lower half because a small business cannot use an API without a developer.
Where it falls short
The terms of service say AssemblyAI may use customer data to train its models, with an opt-out found in its documentation. That default reverses the position of Amberscript and Sonix. The privacy policy says services are mainly hosted in the United States. No editor, no human review.
Wrong for
A buyer who cannot accept training on audio by default, or who has no engineer. Choose Amberscript for a managed European service or Gladia for a developer API with a stated opt-out.
Pricing: Usage-based: Universal-2 at $0.15 per hour and Universal-3.5 Pro at $0.21 pre-recorded; streaming from $0.15 to $0.45 per hour; $50 free credit. (free trial)
Visit AssemblyAI →
#11 Deepgram
Speech-to-text API with per-minute pricing · United States · deepgram.com
API Per minute Real-time
Deepgram prices by the minute, which is rare in a market quoting hours: Nova-3 pre-recorded runs $0.0043 a minute, and the $200 signup credit covers a large test. It also hosts a Whisper Large option, so one account can compare the open model with its own. The privacy policy names Deepgram, Inc. at a San Francisco address and describes processing customer data only on the customer's instructions. For a developer building live captions or a voice feature, it is among the cheaper options, and it fills the last place only because it serves such a narrow buyer.
Where it falls short
The privacy policy states that data is stored on servers in the U.S. and does not address model training, leaving the answer to the DPA. The pricing page did not say whether on-premises deployment is available. Growth starts at $4,000 a year.
Wrong for
A company that needs EU-only storage or a ready-made transcript tool. Amberscript stores in Europe; Speechmatics offers regional and on-premises deployment.
Pricing: Usage-based per minute: Nova-3 monolingual pre-recorded at $0.0043 and streaming at $0.0048 promotional; Growth plan from $4,000 a year; $200 free credit. (free trial)
Visit Deepgram →
What the data says about this market
Of the eleven products ranked here, five are headquartered in the United States: Sonix, Otter, Descript, AssemblyAI and Deepgram. Two are British, Trint and Speechmatics, and one each is Dutch, Irish, French and Singaporean. Seven are finished apps and four are developer APIs. Ten publish at least one price in dollars or euros; Trint is the exception, as its public site shows no plans. Four offer a free plan with no stated end date (Happy Scribe, Notta, Otter and Descript), and the four APIs hand out signup credit, from €50 at Gladia to $200 at Deepgram.
The data terms split the field more sharply than the prices. Amberscript states that its data never leaves Europe and is never used for training. Sonix states it does not train on customer content, and Happy Scribe makes training opt-in. Trint says on its homepage that algorithms are never trained on user transcripts. At the other end, Otter says in its policy that it trains on de-identified audio and transcripts, and AssemblyAI's terms reserve a right to train unless the customer opts out. Two vendors, Deepgram and Notta, left the question open in the pages we read. For a buyer, that makes the privacy policy the most useful document in the evaluation.
A European buyer of speech software is buying across borders. World Bank data puts ICT services at 9.1% of world service exports in 2013 and 14.48% in 2023, and the European Union at 10.58% rising to 17.55%. A transcript is a small piece of that trade: audio recorded in Lyon may be recognized by an engine hosted in Virginia and returned as text within a minute. Several vendors here, among them Amberscript, Trint, Speechmatics and AssemblyAI, now state a European option, and it is worth asking for it in writing before the first file is uploaded.
The 11 ranked vendors, counted
- Headquarters by region: Europe 5, North America 5, Asia-Pacific 1
- By country: United States 5, United Kingdom 2, France 1, Ireland 1, Netherlands 1, Singapore 1
- Pricing model: Usage-based 4, Per user / month 3, Per hour / month 2, Per month, minutes included 1, Per seat / month 1
- Free option: Free trial 5, Free plan 4, None 2
Counted from the 11 vendors on this page. More in our market data.
For the wider market behind transcription software, read our report The Global Shift to ICT Services,
or browse all industry reports.
Questions and answers
What is the best transcription software in 2026?
Amberscript leads this ranking for a small organization buying an app: published euro prices, storage in Europe only and a written promise not to train on your recordings. Happy Scribe is second and the better pick when you want a free plan or 150+ languages, and Sonix is third for occasional users who prefer to pay $10 per hour with no subscription. If you are a developer, skip those and look at Speechmatics or Gladia.
Is automatic transcription accurate enough to use without checking?
For clean audio with one or two speakers, usually close. Happy Scribe says its AI reaches roughly 90 percent on multi-speaker meetings, and Amberscript labels its machine output as 90 percent or better. Names, accents, crosstalk and technical terms are where errors cluster. Read a transcript against the audio before you quote it, and buy human review for anything that will be filed, published or sent to a court.
How much does human transcription cost compared with AI?
Where vendors publish a figure, human-checked transcription starts at €1.75 per minute at Happy Scribe and €1.85 at Amberscript, against machine rates measured in cents per minute. The gap is roughly a factor of ten to fifty. The human tier is justified for legal, medical and published work, where a single wrong word matters, and is wasteful for internal notes.
Which transcription tools store data in the EU?
Amberscript says data is never stored outside Europe. Happy Scribe lists EU data residency on its homepage, and Trint offers EU or US tenants, with EU data in Ireland and Frankfurt. Speechmatics lists EU among its cloud regions, and AssemblyAI offers an EU endpoint at the same price. Sonix states that data is stored outside the EEA and the UK, so it is a poor match for a strict EU requirement.
Do these services use my recordings to train their AI?
It varies by vendor. Amberscript and Sonix say they do not. Happy Scribe uses content for training only if you opt in at signup. Otter's policy says it trains on de-identified audio and transcripts, and AssemblyAI's terms allow training with an opt-out in its documentation. Gladia offers automatic opt-out from its Growth plan. Deepgram's privacy policy does not address the question, so ask before sending sensitive audio.
Can I transcribe a recorded interview without paying?
Yes, within limits. Happy Scribe's free plan includes a ten-minute trial of its AI, Otter gives 300 monthly minutes, and Descript gives 60 media minutes with watermarked export. Notta offers 120 minutes a month but only three minutes per conversation. Sonix gives a 30-minute trial. The APIs hand out credit instead, from €50 at Gladia to $200 at Deepgram.
When should I use an API instead of an app?
When the transcript becomes input to something else: live captions inside your product, call analysis, search over a recording library, or a voice feature. An API costs less per hour, AssemblyAI at $0.15 and Speechmatics from $0.24 on the models listed, but you need code to send audio and store results. For a few interviews a month, an app with an editor is cheaper in total once developer time is counted.
How is transcription software different from an AI meeting assistant?
A meeting assistant joins calls, records them and writes a summary with action items, and its price assumes a seat per person. Transcription software starts from a file or a stream and returns text, charged by the hour or minute. Otter, Notta and Happy Scribe straddle both. If most of your audio is calls, our ranking of AI meeting assistants is the better starting point.
What file formats and languages can I expect?
Most apps take common audio and video files and export text, subtitles and captions; the specific format list is on each vendor's help pages, which we did not audit. Language range does differ: Happy Scribe names 150+, Amberscript 90+, Speechmatics 55+, Sonix 54+, Notta 58 and Trint 50+ for live transcription. Check your language and accent against the sample before paying.
Is Otter the best choice because everyone uses it?
It is the best known, which is why it is marked as a dominant vendor and kept out of the top three. Its free plan is usable for live meetings, but file imports are limited to three for life and the policy allows training on de-identified recordings. For client or research audio, Amberscript, Happy Scribe or Sonix give clearer data terms.
What should I check before uploading client or patient audio?
Four things: whether the vendor will sign a data processing agreement, which region stores the audio, whether the default is training on your content, and how deletion works. Sonix, Speechmatics, Gladia and Notta list SOC 2 or HIPAA claims; Amberscript lists ISO 27001. Certification on a homepage is not a contract, so ask for the terms in writing.
