The two-minute version

Two of the twenty-two tools here carry a word error rate that a third-party benchmark produced. A couple more carry accuracy figures that came from reviewers rather than the vendor. The rest are asking you to take their word for it.

AssemblyAI's Universal-3.5 Pro led all 31 models on Coval's independent July 2026 dataset at 3.4% word error rate (WER, the share of words a system gets wrong). Speechmatics posts a 1.07% WER from Pipecat's August 2026 benchmark of 12 speech-to-text services, the lowest of the twelve, and Speechmatics' own site is where you read that result. Same class of test, and both of them reach you through the vendor's own page. Neither one hands you a permanent free tier, so if your budget is zero the answer changes shape: self-hosted Whisper is free under an MIT license and covers developer workloads, and Fathom's free plan records and transcribes meetings with no cap.

Here's the pattern worth carrying through the rest of this page. Between the raw developer APIs and the meeting notetakers, the free plans get more generous and the accuracy evidence gets softer. Count it. Four of the seven developer APIs here have no permanent free plan at all, only starter credits, while every one of the six meeting notetakers does. Metered APIs publish per-hour rates and, sometimes, a word error rate someone else measured. Notetakers publish free plans, and the accuracy numbers they do publish are their own. Both get sold under the same keyword, and they are two different purchases.

Five shelves, one keyword

Work out which shelf you're shopping before you compare a single price. Most of this list is irrelevant to you, and that's good news.

Developer APIs

Speechmatics (1), Google Cloud Speech-to-Text (4), AssemblyAI (5), OpenAI Whisper (7), Gladia (8), Deepgram (11), ElevenLabs Scribe (15). Metered by the minute or hour, wired in with code, and the shelf where accuracy tends to arrive as a word error rate rather than a headline percentage.

Meeting notetakers

Fathom (2), Fireflies.ai (3), Otter.ai (6), Granola (9), Notta (12), tl;dv (22). Seat-priced, calendar-aware, and sold on summaries and CRM sync rather than raw transcript quality.

Production and editing suites

Sonix (10), Descript (14), Happy Scribe (16), TurboScribe (19), Trint (21). Built for media work: captions, subtitles, editors, exports.

Dictation layers

Wispr Flow (13), Superwhisper (18). They live in your keyboard, not your meeting invite.

AI plus human hybrids

Rev (17), Verbit (20). Machine transcription with a paid human pass behind it, for legal, accessibility and compliance work where a wrong word is a real problem.

Quick picks

Two markets, so two shortlists. This one is for the engine shelf, where your decision comes down to latency, cost per hour, and what the contract includes before a salesperson gets involved.

Best third-party benchmark numbers: Speechmatics at 1.07% WER, lowest of the 12 services in Pipecat's August 2026 run, and AssemblyAI at 3.4%, first of 31 models on Coval's July 2026 dataset. Different datasets, so not a head-to-head.
Cheapest hour of audio: self-hosted Whisper at $0 under an MIT license. Among metered APIs, AssemblyAI's $0.15/hour is the lowest published hourly rate; Speechmatics quotes $0.129 per unit, which won't sit on that axis without a conversion of your own.
Compliance without a contract: AssemblyAI, which lists a HIPAA BAA, PCI-DSS, SOC 2 Type 2 and ISO 27001 at no premium, with the EU region priced the same as U.S.
Lowest published real-time latency: ElevenLabs Scribe v2 Realtime, under 150 ms, vendor-stated. Gladia's sub-300ms is the EU-headquartered alternative.
Run it yourself: Whisper, MIT-licensed across six model sizes, or Speechmatics' on-premises virtual appliance.

Standing at the other shelf instead? Fathom's free plan doesn't meter your recordings or transcriptions, and Fireflies, Fathom, Notta and Granola all publish enough compliance paperwork to clear a procurement review.

All 22 at a glance

Tool	Shape	Entry price	$ per hour of audio	Free access
1. Speechmatics	Developer API	$0.129 per unit	$0.129 per unit, priced per unit rather than per hour	$100 credit, no card, 2 concurrent real-time sessions
2. Fathom	Notetaker	Free, then $20/mo	Seat-priced, not metered	Unlimited recordings and transcriptions
3. Fireflies.ai	Notetaker	$10/seat/mo annual	Seat-priced, not metered	400 min storage, 2-hr meetings
4. Google Cloud Speech-to-Text	Developer API	$0.004 to $0.024 per minute	$0.96 with data logging, $1.44 without, falling to $0.24 above 2,000,000 minutes/month	60 min/mo on V1, plus up to $300 in new-customer credits
5. AssemblyAI	Developer API	$0.15/hour	$0.15	$50 credit, no card
6. Otter.ai	Notetaker	Free, then $16.99/user/mo	Seat-priced, with a monthly minute allowance per seat	300 min/mo
7. OpenAI Whisper	Open model plus API	$0.003 to $0.006 per minute	$0.36 hosted, $0.18 on mini, $0 self-hosted	MIT license, six model sizes, no usage cost
8. Gladia	Developer API	$0.61/hour async	$0.61, from $0.20 on Growth	€50 trial credit, no expiry stated
9. Granola	Notetaker	Free, then $14/user/mo	Seat-priced, not metered	30-day meeting history
10. Sonix	Production suite	$10/hour, or $25/mo	$10	30-minute trial
11. Deepgram	Developer API	$0.0048/min streaming	$0.29	$200 credit, no card, no expiry
12. Notta	Notetaker	Free, then $8.17/mo annual	Seat-priced, with a monthly minute allowance per seat	120 min/mo, 3 min per conversation, 1 seat
13. Wispr Flow	Dictation	Free, then $15/user/mo	Seat-priced, not metered	2,000 words/week desktop, 1,000/week iPhone, unlimited Android
14. Descript	Production suite	Free, then $16/mo annual	Seat-priced, not metered	60 min of media/mo, 5GB, 720p watermarked export
15. ElevenLabs Scribe	Developer API	Free, then $6/mo	Credit-priced, roughly 330 credits per minute	10,000 credits/mo, about 30 minutes of audio
16. Happy Scribe	Production suite	$8.50/mo annual	Allowance-priced: 120 AI minutes on the $8.50 plan, extra minutes $0.20/min	10 AI minutes, 45-min meeting duration per recording
17. Rev	AI plus human	Free, then $25.49/mo annual	Seat-priced, with 5,000 AI minutes per seat; human pass $1.99/min	45 AI min/mo, English only, 1 seat
18. Superwhisper	Dictation	$8.49/mo	Seat-priced, not metered	3,000 words of Pro trial, then core features unlimited
19. TurboScribe	Transcriber	$10/mo annual	Seat-priced, not metered	3 transcripts/day, up to 30 min each
20. Verbit	AI plus human	From $24/mo for 100 hours	$0.24 across the entry plan's 100 hours	No free tier and no trial
21. Trint	Production suite	The self-serve pricing pages do not load, so there is no tier price to read	-	-
22. tl;dv	Notetaker	€29/mo per seat	Seat-priced, not metered	A €0 plan is offered

Read the fourth column twice. It's the same hour of audio, priced by each vendor in its own unit, converted here so you can line them up on one axis. One row resists that: Speechmatics, the top-ranked tool on this page, publishes $0.129 per unit plus volume discounts rather than an hourly rate, so lining it up against an hourly competitor takes a conversion step of your own.

The accuracy-evidence ledger

Is the feature list what separates these tools? No. It's the evidence class. Read the middle column, then read what the tool's own reviewers say breaks.

Tool	Accuracy claim	Who checked it	Reported failure mode
Speechmatics	1.07% WER, lowest of 12 services tested	Pipecat benchmark, August 2026, published on Speechmatics' own site	Slow processing and latency in real-time use
AssemblyAI	4.35% normalized WER pre-recorded, 5.53% streaming	Vendor's own; separately 3.4%, first of 31 models, on Coval's independent July 2026 dataset, published on AssemblyAI's own site	Heavy noise, overlapping speakers, strong accents
Google Cloud	Chirp 3, trained on millions of hours and 28 billion sentences	Vendor's own; G2 4.5/5 from 318 reviews	Two-month average implementation time
Deepgram	54.2% WER reduction with Nova-3 streaming	Vendor's own	Accents, noisy audio, niche terminology
Gladia	9.6% WER on real English audio (Solaria-3)	Vendor's own; G2 reviewers praise accented speech	Occasional service downtime
ElevenLabs Scribe	WER tiers by language: 5% or better for 34 languages	Vendor's own, published per language	18 languages sit in a 25% to 50% WER tier
OpenAI Whisper	98 languages on whisper-1	Vendor docs; G2 reviewers report minor edits only	Hallucinated text on unclear or non-speech audio
Sonix	99% across 54+ languages	Vendor's own; Capterra reviewers report 98% to 99%	Background noise, accents, overlapping speakers
Notta	98%+ across 58 languages	Vendor's own	Multiple speakers, accents, noisy audio
Fireflies.ai	95% across 100+ languages	Vendor's own	Heavy accents, overlapping speech, background noise
Happy Scribe	Up to 99% on clear audio, 94% diarization	Vendor's own	File conversion producing unusable output
Descript	Up to 95% with clear audio	Vendor's own docs	Files over 15 hours fail to auto-transcribe
Rev	47% more accurate than competitors in challenging environments	Vendor's own	Foreign languages, 3+ hour multi-speaker audio
Trint	40+ languages with automatic detection	Vendor's own	French and German accuracy complaints from reviewers
TurboScribe	Powered by Whisper, three accuracy modes	Vendor's own; Trustpilot 4.4/5 from 430 reviews	Processing failures on large or long files
Wispr Flow	Clean, filler-free dictation	G2 reviewers report 120 to 250 WPM	Hindi and Hinglish called barely functional
Fathom	Marketing adjective	G2 5.0/5 across roughly 6,942 reviews	Isolated accuracy misses, call-recording issues
Otter.ai	Real-time transcription with speaker recognition	G2 4.4/5 from 442 reviews	Accents, background noise, speaker ID gaps
Granola	Background transcription of computer audio	G2 4.7/5 from 34 reviews	Imperfect speaker and voice distinction
Superwhisper	Whisper Large plus context-aware Super Mode	Vendor's own	Contextual answers instead of literal transcription
Verbit	Captivate ASR (automatic speech recognition) plus human QA	Vendor's own	-
tl;dv	Marketing adjective	Vendor's own	-

Notice how often "vendor's own" repeats, and notice which failure mode recurs where one is named at all: accents, crosstalk and background noise account for six of these rows, spread across the API, notetaker and production shelves, while the dictation shelf shows none of the three. That's the honest state of the category.

What "free" actually buys you

The same trick runs on the pricing pages. A free plan and a trial dressed as one are different products, and the free column above gives you the cap wherever the vendor publishes one. Two rows don't: tl;dv confirms only that a €0 plan exists, and Trint's self-serve pages don't load at all. Rule of thumb: a starter credit is a balance, not a plan. Four of the seven developer APIs here hand you a pile of dollars instead of a free tier, though Deepgram's $200 carries no expiration and Gladia's €50 states none. Sonix and Happy Scribe give you 30 minutes and 10 minutes, Verbit gives you nothing, and the uncapped free offers include Fathom, self-hosted Whisper, Superwhisper's core features after its 3,000-word Pro trial, Wispr Flow on Android, and Fireflies, which runs unlimited transcription and summaries on every tier while capping storage at 400 minutes.

The ranking

1. Speechmatics: the checked number at the top

Shape: developer API, deployable on cloud, Docker/Kubernetes, or an on-premises virtual appliance
Price: $0.129 per unit on Pro, with a 20% discount available and volume discounts published on the pricing page. Enterprise is custom and scaled to volume
Free: $100 in credit, no card required, 2 concurrent real-time sessions
Accuracy evidence: 1.07% WER in Pipecat's August 2026 benchmark, lowest of 12 services, published on Speechmatics' own site
Wires into: LiveKit Agents, Pipecat, Vapi

That one number is why Speechmatics leads, with a qualification attached. Pipecat ran the benchmark. Speechmatics' own site is where the result lives. Only one other tool on this page carries a third-party word error rate at all, and most carry none, so the qualification trims the claim without erasing it. The company has been at this since 2006, when Dr. Tony Robinson founded it as Cantab Research Ltd., and it puts its total funding at $90.6M, including a $62M Series B in 2022.

The feature surface goes well past raw transcription: diarization, custom dictionaries, translation, summarization, chapter and topic detection, language ID, plus text-to-speech and voice-agent APIs. G2 reviewers score it 9.4/10 for ease of use and 97% on installation, and hand it a Best Support badge on a 9.1/10 support score. If you're already building on a voice-agent framework, your integration cost is close to nothing, because the paths into LiveKit Agents, Pipecat and Vapi are first-party.

There's no permanent free tier, reviewers call the pricing expensive next to alternatives, and the rate arrives per unit rather than per hour, so setting it beside an hourly competitor takes a conversion step of your own. Some reviewers also report slow processing that bites in real-time scenarios. So the adoption cost here is a pricing conversation before the pilot, which is fine if you need on-prem deployment and a defensible accuracy story in one purchase, and painful if you wanted to swipe a card and start.

2. Fathom, the free plan that doesn't run out

Shape: meeting notetaker for Zoom, Google Meet and Microsoft Teams
Price: Premium $20/month ($16 billed annually), Team $19/user/month ($15 annually, minimum 2 users), Business $34/user/month ($25 annually), Enterprise custom
Free: unlimited recordings and transcriptions, free forever
Accuracy evidence: vendor's own language, alongside a 5.0/5 G2 rating across roughly 6,942 reviews
Wires into: Salesforce, HubSpot, Slack, plus a public API and an MCP server

The free plan doesn't meter you: recordings and transcriptions are unlimited, free forever, on the published plan. Against Notta's 3 minutes per conversation and Rev's 45 minutes a month, that's a genuinely different offer, and it means your evaluation can run as long as your patience does.

The paid ladder adds AI action items and a conversational assistant at Premium, cross-call search and collaboration at Team, then CRM field sync, Deal View and AI Scorecards at Business. Security is solid for the shelf on Fathom's own documentation: SOC 2 Type II audited, HIPAA compliant, GDPR-aligned, certified under the EU-U.S., UK and Swiss Data Privacy Frameworks, with an explicit opt-out from de-identified data use in model training. The same site puts the saving at 6+ hours weekly on follow-up and about 38 minutes per meeting.

What it costs you to adopt is mostly patience with two gaps. Reviewers report occasional call-recording and meeting-management trouble, and SSO, SCIM and custom retention all sit behind custom Enterprise pricing, so a security review will end in a sales call even though your users never had to. There's no independent accuracy benchmark to point at either, which is normal on this shelf and still worth knowing.

3. Fireflies.ai, the notetaker built like enterprise software

Shape: notetaker plus conversation intelligence
Price: Pro $10/seat/month annually ($18 monthly), Business $19/seat/month annually ($29 monthly), Enterprise $39/seat/month annual only
Free: 400 minutes of team storage, 2-hour recording limit per meeting
Accuracy evidence: 95% across 100+ languages, vendor's own
Wires into: 100+ apps including Salesforce, HubSpot and Zapier, plus API access

Fireflies behaves like it expects a procurement review, and on this shelf it has company in Fathom, Notta and Granola. The security page states SOC 2 Type II, GDPR and HIPAA compliance, a 0-day data retention policy with vendors and partners, 256-bit AES encryption at rest with TLS in transit, and a default position of not training on customer data. Enterprise adds SSO/SCIM, HIPAA/FERPA and custom retention. Unlimited transcription and summaries run on every tier.

On top sits an unusually deep AI layer for a notetaker: AskFred for meeting Q&A, Live Assist for real-time coaching, an Email Assistant, 200+ industry-specific AI Skills, Voice Agents, and conversation intelligence with talk-time, sentiment and topic trackers. Sentiment across 1,481+ reviews is strong, with Capterra at 4.9/5 from 735 reviews.

Your real cost is tier drift. API access, analytics, unlimited storage and SSO all sit above the entry plan, so the $10 seat you budget for is rarely the seat you end up buying. The complaints are consistent too: accuracy drops with heavy accents, overlapping speech and background noise, support is email and chat only with no phone line, and some users report billing confusion. Keep in mind that the 95% headline is a marketing figure, not a benchmark result.

4. Google Cloud Speech-to-Text: power with a pricing matrix attached

Shape: developer API inside Google Cloud
Price: V1 Standard $0.016/min with data logging or $0.024/min without, after 60 free minutes/month; V1 Medical $0.078/min; V2 Standard $0.016/min falling to $0.004/min above 2,000,000 minutes/month; V2 dynamic batch $0.003/min
Free: 60 minutes/month on the V1 API, plus up to $300 in credits for new Google Cloud customers
Accuracy evidence: Chirp 3, trained with self-supervised learning on millions of hours of audio and 28 billion sentences across 100+ languages, vendor's own; G2 rates the product 4.5/5 from 318 reviews
Wires into: Vertex AI SDK, Dialogflow, Cloud Storage, and 13 more verified integrations listed on G2

Read that price split twice. You pay less per minute if you let Google log your data, which tells you something useful about the default. Elsewhere in the trust column, Google documents V2 data residency in regions like Singapore and Belgium, customer-managed encryption keys, audit logging and an on-prem deployment option.

Capability is broad: synchronous recognition up to 1 minute, asynchronous up to 480 minutes, streaming in real time, speaker diarization, multichannel recognition, model adaptation for domain vocabulary, and specialized models for phone_call, video, telephony, medical_dictation and medical_conversation. Support runs to 85+ languages and variants.

So what stops it climbing higher? Time and complexity. G2's aggregate time-to-implement is 2 months, and the price is a matrix of API version, model SKU, data-logging choice, channel count and volume bracket rather than a rate you can quote from memory. If your stack already lives on Google Cloud, that friction is mostly amortized. If it doesn't, the honest cost of adopting this is a round of cost modeling before the pilot even starts.

5. AssemblyAI, the compliance you don't have to negotiate for

Shape: developer API plus voice-agent platform
Price: $0.15/hour pre-recorded on Universal-2, $0.21/hour on Universal-3.5 Pro, $0.15/hour streaming, $4.50/hour ($0.075/min) for the Voice Agent API
Free: $50 in credits, no card required
Accuracy evidence: 4.35% normalized WER pre-recorded and 5.53% streaming on its own benchmarks, against 5.338% for OpenAI GPT-4o Transcribe and 6.662% for Deepgram Nova-3; separately, 3.4% WER and first place among 31 models on Coval's independent July 2026 dataset
Wires into: LiveKit, Pipecat, plus Python and JavaScript/Node.js SDKs

Here's the line that matters if you work anywhere near regulated data. AssemblyAI lists a HIPAA BAA (business associate agreement), PCI-DSS, SOC 2 Type 2 and ISO 27001 as included at no premium, and prices the EU region with GDPR-compliant residency the same as U.S. No enterprise contract, no salesperson, no procurement detour. That combination is rare enough on this list to decide the purchase on its own.

The platform spans pre-recorded and real-time transcription, a Voice Agent API that runs speech to speech over one WebSocket, Speech Understanding for diarization, summarization and sentiment, Guardrails for PII (personally identifiable information) redaction and content moderation, and an LLM Gateway. G2 reviewers put ease of use at 9.3/10 and quality of support at 9.6/10, and named it a Leader in the Fall 2025 Voice Recognition Grid. The company puts its funding at roughly $115M.

Weak spots are honest ones. There's no permanent free tier, users report costs climbing quickly at high audio volume, and accuracy and latency degrade with heavy background noise, overlapping speakers or strong regional accents. Familiar failure modes, and better documented than on the tools that publish no benchmark at all. Budget your adoption in engineering time rather than legal review, which is the reverse of most of this list.

6. Otter.ai

Shape: meeting notetaker
Price: Pro $16.99/user/month ($8.33 annually), Business $30/user/month ($19.99 annually), Enterprise custom
Free: Basic at 300 transcription minutes/month
Accuracy evidence: G2 4.4/5 from 442 reviews with a 9.0 ease-of-use sub-score
Wires into: Zoom, Salesforce, HubSpot, plus Slack, Notion, Asana, ClickUp, Dropbox, Google Drive, Zapier and Claude via MCP

Start with the compliance posture, because for some of you it ends the conversation. Otter's security page carries no SOC 2 or GDPR certification claim, and HIPAA is sold as an add-on to the custom-priced Enterprise plan. For a regulated workload that's a different starting position than AssemblyAI, which lists the paperwork on its pricing page, or Fireflies, which lists it on a security page you can read without a call.

Then the minutes. 300 a month free, 1,200 on Pro with a 90-minute meeting ceiling, and unlimited meetings only at $30/user/month Business, which allows 4-hour meetings. Add up your team's real monthly minutes before you pick, because the jump from Pro to Business is where your actual cost lives.

What Otter has is reach and polish. Real-time transcription with speaker recognition, AI Chat across meetings, OtterPilot guidance and a Meeting Agent notetaker, with reviewers consistently praising setup and transcript quality. They also report gaps in speaker identification and summary completeness, plus the familiar trouble with accents and background noise.

7. OpenAI Whisper, the zero-dollar floor under this whole market

Shape: open-weight model plus a hosted API
Price: whisper-1 $0.006/minute, gpt-4o-transcribe $0.006/minute, gpt-4o-mini-transcribe $0.003/minute, gpt-4o-transcribe-diarize $0.006/minute
Free: the open-source model under an MIT license, self-hosted at no usage cost
Accuracy evidence: 98 languages on whisper-1, vendor docs; G2 reviewers report transcripts usually needing only minor edits, at 4.5/5 from 15 reviews
Wires into: REST endpoints for transcription and translation, plus an open-source pip package for any ML pipeline

Every per-minute price on this page is being quoted against a free alternative, and this is it. Six model sizes ship from tiny at 39M parameters to large at 1550M, plus an 809M turbo variant that runs about 8x faster than large. Run it on your own hardware and your audio never leaves the building, the same answer Speechmatics' on-premises appliance and Superwhisper's local processing give you. On the hosted side, OpenAI states API inputs and outputs are not used for training by default and are retained up to 30 days for abuse monitoring.

What does free cost you? Work. Reviewers say local setup feels technical, performance drags on large files and real-time use, and support quality is inconsistent. The classic whisper-1 model has no native speaker diarization, though gpt-4o-transcribe-diarize adds it with up to 4 reference speaker clips. It can hallucinate text on unclear or non-speech segments, and the hosted API caps uploads at 25 MB per file across mp3, mp4, mpeg, mpga, m4a, wav and webm, so long recordings need chunking.

Tip: if you're pricing a high-volume pipeline, benchmark self-hosted Whisper first. It sets the number every vendor quote has to beat.

8. Gladia, an EU-native API with a real-time bias

Shape: developer API for voice AI
Price: Starter $0.61/hour async and $0.75/hour real-time; Growth from $0.20/hour async and $0.25/hour real-time; Enterprise custom
Free: €50 in one-time trial credits, no ongoing free plan
Accuracy evidence: 9.6% WER on real English audio from Solaria-3, vendor's own, with G2 reviewers independently praising accuracy on accented speech
Wires into: Pipecat, LiveKit, Twilio, plus Vapi, Recall, Zapier, Make and n8n

Gladia is EU-headquartered and sells that position hard, with SOC 2 Type II, GDPR, HIPAA and ISO 27001 compliance stated on the pricing page, a zero-data-retention option on Enterprise and automatic model-training opt-out on Growth. If data sovereignty is the gate you have to clear, that shortens your conversation considerably.

The engine claims sub-300ms latency across 100+ languages, and the API bundles diarization, any-to-any translation in the same call, code-switching and audio intelligence including named entity recognition, sentiment, summarization and PII redaction. Gladia says it has processed 2+ billion minutes for 300,000+ developers. Your adoption cost is the jump from Starter to Growth, since the headline $0.20 rate arrives only with volume you have to commit to.

The caution is thin evidence rather than bad evidence. G2 shows 4.8/5 across just 23 verified reviews, one of which flags unclear pricing at enterprise volume, another occasional downtime and difficulty tracking usage at scale. Small sample, strong signal, and worth your own load test before you commit.

9. Granola

Shape: meeting notetaker that captures computer audio directly
Price: Business $14/user/month, Enterprise $35/user/month
Free: Basic at $0, with meeting history limited to 30 days
Accuracy evidence: G2 4.7/5 from 34 reviews, down from an all-five-star start at 11 reviews
Wires into: Attio, Notion, Slack, plus HubSpot, Affinity, Zapier, MCP and the API, all from the $14 tier

Nothing joins your call. That is the draw here, and it's a social one. Granola transcribes computer audio in the background across Zoom, Google Meet and Teams, then turns your own partial notes into structured ones. Reviewers describe setup as extremely smooth and specifically value not putting a bot in front of a client.

Privacy is a genuine strength on the company's own documentation. Granola publishes SOC 2 Type 2 verification by independent auditors, a GDPR Data Processing Agreement, and no third-party model training on user data. No audio is retained at all, only transcripts and notes, held in a U.S.-hosted AWS VPC. Notes stay private by default, and there's a public vulnerability disclosure policy with post-mortems.

Where does it bite? The free tier's 30-day history means your searchable past evaporates on a rolling basis, so the moment Granola becomes useful to you is the moment it costs $14 a seat. Reviewers also report occasional transcript inaccuracy and imperfect speaker distinction.

10. Sonix: the transcription workbench

Shape: production and editing suite
Price: Pay As You Go $10/hour prorated to the second; Core $25/month or $275/year for 5 hrs/month; Advanced $50/month or $550/year for 20 hrs/month; Pro $80/month or $880/year for 40 hrs/month; Enterprise custom
Free: a 30-minute trial, not an ongoing plan
Accuracy evidence: 99% across 54+ languages, Sonix's own claim, with Capterra reviewers reporting 98% to 99% in practice at 4.9/5 from 143 reviews
Wires into: Zoom, Adobe Premiere, Zapier, plus a REST API with webhooks

Buy Sonix when the transcript itself is the deliverable. Neural translation covers 55+ languages, Sonix bills its automated captions and subtitles as frame-perfect, and the in-browser editor keeps text synced to audio while AI analysis produces summaries, chapters and sentiment through structured conversation objects. Trust holds up too on the company's own documentation, with SOC 2 Type II certification, HIPAA compliance, AES-256 encryption in transit and at rest, and a zero-training policy on customer data. That 99% figure is the vendor's, and the Capterra reports of 98% to 99% in practice are the closest thing to corroboration this shelf offers.

The pricing model is where it will annoy you. Reviewers repeatedly flag mandatory monthly fees during quiet months and totals that are hard to predict, and there's no permanent free tier to soften the entry. Integrations stay inside the media and meeting world, with no CRM and no broad app ecosystem. If your volume is lumpy, model the $10/hour pay-as-you-go line first and ignore the subscriptions entirely.

11. Deepgram, low latency and a commitment cliff

Shape: developer API for real-time voice
Price: Pay As You Go usage-based, for example $0.0048/minute for Nova-3 streaming, which is $0.29 an hour; Growth at $4K+/year in pre-paid credits saving up to 20%; Enterprise custom
Free: $200 in credit, no card, no expiration, but no permanent free tier
Accuracy evidence: a 54.2% WER reduction claim for Nova-3 streaming, vendor's own
Wires into: Twilio, Cloudflare, Vapi

Deepgram is built for production voice applications and the reviews read that way: fast, low-latency performance, a developer-friendly API that goes in quickly, and responsive support. G2 puts ease of use at 9.1/10 and support quality at 8.8/10 across 326 reviews, with 91% saying the company is headed in the right direction.

The model lineup is unusually wide. Flux and Nova-3 handle streaming, Nova-2 ships 11 domain-specific variants, Aura-2 and Flux TTS cover synthesis, and a combined Voice Agent API stitches it together. Nova-3 supports 60+ languages; Flux supports 10 including Spanish, French, German, Hindi and Japanese. SOC 2 Type II certification comes from independent auditor Cyberguard Compliance, with HIPAA BAA availability and GDPR readiness stated in the docs.

So what happens when the $200 runs out? Pay As You Go keeps metering at $0.29 an hour with no minimum and no expiration. The next tier up is a $4K commitment or a sales call, and that step is the moment your pilot becomes a budget line someone has to defend. Reviewers also report accuracy varying on accents, noisy audio and niche terminology, and find the model and pricing options complex at first contact.

12. Notta

Shape: meeting notetaker and transcription app
Price: Pro $8.17/month billed annually ($97.99/year) or $13.99 monthly; Business $16.67/month annually ($199.99/year) or $27.99 monthly; Enterprise custom from 51 seats
Free: 120 transcription minutes/month, 3 minutes maximum per conversation, 1 seat
Accuracy evidence: 98%+ across 58 languages, vendor's own
Wires into: Zoom, Salesforce, HubSpot, plus 21 more native connections and Zapier for 5,000+ apps

Look at that free tier again. It's 120 minutes a month, capped at 3 minutes per conversation. Almost no real meeting fits inside that, which is exactly what reviewers complain about. Treat it as a demo, not a runway.

Paid, Notta is genuinely capable. Pro lifts you to 1,800 minutes a month with recordings up to 5 hours, Business removes the minute cap, and the product goes past transcripts into real-time translation, AI summaries and action items, Notta Brain for cross-meeting search and Q&A, and generated presentations, infographics and Word, PDF or Excel documents. The compliance write-up is one of the more detailed on this shelf: SOC 2 Type II, ISO 27001, GDPR and CCPA, HIPAA guideline adherence, AES-256 at rest, TLS 1.2 in transit. Two things will cost you: no public developer API is documented, so deeper automation routes through Zapier, and the 98%+ number is Notta's own, with reviewers reporting the usual degradation on multiple speakers, accents and noisy rooms.

13. Wispr Flow, dictation that keeps up with your mouth

Shape: dictation layer across Mac, Windows, iPhone and Android
Price: Pro $15/user/month ($12 annually); Growth/Enterprise from $18/user/month annually (from $23 monthly)
Free: 2,000 words/week on desktop, 1,000/week on iPhone, unlimited on Android
Accuracy evidence: G2 reviewers report typing speeds moving from 50 to 60 WPM up to 120 to 250 WPM, with business-specific terminology captured correctly
Wires into: Claude and ChatGPT via MCP, plus Calendar and Slack, all available on the free plan

This one isn't competing with the APIs. It works in any text field, turns rambling speech into clean formatted text, and gets out of the way. The personalization is the good part: a personal or team dictionary, text snippets, tone adapters per app, a notetaker with speaker ID, and recognition of developer vocabulary like Supabase, Cloudflare and Vercel. Compliance is stated up front rather than gated, with SOC 2 Type II, ISO 27001 and HIPAA-readiness with a BAA, and parent company Wispr reports raising $81 million.

Three honest dents. Mobile is weaker than desktop, with keyboard-switching and permission friction. Dictation reportedly fails after long sessions. And the 100+ languages headline doesn't hold evenly, with reviewers calling Hindi and Hinglish barely functional. One more thing to weigh before you roll this out to a team: Wispr Flow's Trustpilot business profile has been removed for violating Trustpilot's guidelines, which the official site doesn't address.

14. Descript: transcription as an editing interface

Shape: production suite where the transcript is the timeline
Price: Hobbyist $24/month ($16 annually), Creator $35/month ($24 annually), Business $65/month ($50 annually), Enterprise custom; White Glove human review at $2.00/minute
Free: 60 minutes of media/month, 5GB storage, 100 one-time AI credits, 720p export with watermark
Accuracy evidence: up to 95% with clear audio, Descript's own docs, with the paid human-reviewed tier advertised at up to 99% in about 24 hours
Wires into: Premiere, Final Cut, DaVinci Resolve, plus Pro Tools, Zoom, Slack, HubSpot, an official API and MCP

Across 1,103 reviews, stability is one of the recurring complaints. Users report lost edits, corrupted projects and crashes on larger projects, support that's slow and largely bot-driven, and confusing non-prorated upgrades. Files over 15 hours will not auto-transcribe. Value for money sits at a middling 4.2/5, and that's the number to sit with before you move a client project into it.

The pitch, when it holds, is excellent: cut a sentence by deleting the words. Podcasters and video editors love it, with Capterra reviewers scoring ease of use at 4.4/5 across 183 reviews and repeatedly calling the text-based workflow a time-saver. Around it sit multilingual transcription across 22 languages beyond English, an AI Speaker Detective for diarization, filler-word removal, Studio Sound cleanup and captions, with SOC 2 Type II, SAML SSO/SCIM and GDPR compliance stated. If you want a speech-to-text engine rather than an editor, this isn't your shelf.

15. ElevenLabs Scribe, the vendor that publishes its bad languages too

Shape: developer API inside a broader voice-AI platform
Price: Starter $6/month for 30,000 credits, Creator $11/month ($22 first month) for 121,000, Pro $99 for 600,000, Scale $299 for 1.8M, Business $990 for 6M; speech to text runs about 330 credits per minute
Free: 10,000 credits/month
Accuracy evidence: published WER tiers by language, vendor's own, with 34 languages at 5% or better out of 90+ supported, and 18 languages in a 25% to 50% tier
Wires into: REST plus WebSocket streaming, with first-party Python and Node.js SDKs

Give credit where it's earned. Scribe is the only tool here that tells you accuracy language by language instead of selling you a language count. Eighteen languages sitting at 25% to 50% WER is not a flattering disclosure, and publishing it is the behavior this category needs more of. If your workload is Swahili or Tamil rather than English, that table is worth more than any headline percentage on this page. Feature depth is serious too: speaker diarization up to 32 speakers, entity detection across 65 entity types, keyterm prompting up to 1,000 terms, word-level timestamps, dynamic audio tagging and PII redaction, with ElevenLabs putting Scribe v2 Realtime latency under 150 ms on its speech-to-text page.

The pricing is the sore spot. How many minutes is 10,000 free credits? About thirty. Credits are shared across text to speech, speech to text and dubbing, with no speech-to-text minute allowance per tier, so you back-calculate your own usage every month. Users on G2 call the credit system restrictive and confusing, and note there's no preview before credits are spent. The 4.5/5 rating from 1,140 reviews covers the whole platform rather than Scribe specifically.

16. Happy Scribe, wide language reach with a human backstop

Shape: production suite plus optional human transcription
Price: Basic $8.50/month annually ($17 monthly) for 120 AI minutes, Pro $19/month annually ($29 monthly) for 600 minutes, Business $59/month annually ($89 monthly) for 6,000 minutes and 5 seats, Enterprise on request; extra minutes at $0.20/minute; human transcription and subtitling from $1.90 to $2.00 per minute across 65+ languages
Free: a 10-minute AI trial with a 45-minute meeting duration per recording
Accuracy evidence: up to 99% on clear audio and 94% speaker diarization, vendor's own
Wires into: Google Meet, Teams, Zoom, plus third-party models including Claude, ChatGPT and Mistral on paid plans

Happy Scribe says it covers 150+ languages, the highest figure any vendor on this page publishes, and pairs machine output with a human-made option when a client needs the transcript to be right rather than fast. An AI notetaker joins your calls, subtitles generate automatically with human refinement available, and the company documents SOC 2 Type II certification, GDPR compliance, EU data residency and opt-in rather than default AI training. Both accuracy figures above are its own too.

Is ten minutes enough to evaluate this? No, so plan on a paid month. Trustpilot sentiment is good at 4.7/5 across 1,398 reviews, with speed and a straightforward interface as the recurring praise, but the complaints cluster in two places: file conversion producing broken or unusable output, and transparency around registration, trial terms and export limits, with the free offer described by some users as misleading.

17. Rev

Shape: hybrid AI plus human transcription and captioning
Price: Essentials $25.49/month billed annually ($29.99 monthly) for 5,000 AI minutes/seat and up to 3 seats; Pro $47.99/month annually ($59.99 monthly) for 10,000 minutes/seat, 5 seats and the Clio integration; Unlimited is custom; human transcription at $1.99/minute; Global Subtitles from $6.49 to $15.99 per minute across 17 languages
Free: 45 AI minutes/month, English only, 1 seat
Accuracy evidence: "47% more accurate than competitors in challenging environments", vendor's own, on a model trained on 50,000+ hours and supporting 57+ languages
Wires into: Clio, plus REST APIs with Python and Node SDKs

Rev has been at this since 2010 and reports a marketplace of 72,000+ human transcription and caption contributors behind the software. That's the real product. When the AI pass isn't good enough for a deposition or a broadcast caption, you buy a person for $1.99 a minute without changing tools, a verification pass Rev puts at 99%+ accuracy. Capterra reviewers rate it 4.6/5 across 51 reviews and describe the interface as no fuss. For developers, docs.rev.ai covers fully punctuated context-aware transcripts, inverse text normalization, word-level timestamps and custom vocabulary.

Where it struggles is written on the tin. Reviewers report weak foreign-language support despite the 57+ language claim, and difficulty with very long or complex multi-speaker audio. Native integrations are thin, with Clio the only connector named on the pricing page, and CJIS/HIPAA compliance plus SSO arrive only on the custom Unlimited tier. Your budget has to carry two stacks at once: seats for the machine, per-minute for the humans.

18. Superwhisper: maximum control, minimum track record

Shape: dictation app for macOS, Windows and iOS
Price: Pro $8.49/month with monthly, yearly or lifetime billing, a 40% student discount and a 30-day refund guarantee; Enterprise custom
Free: 3,000 words of Pro-feature trial, then unlimited use of core features including Whisper transcription, 100+ languages and meeting recording
Accuracy evidence: vendor's own, built on Whisper Large with a context-aware Super Mode
Wires into: Slack, Gmail, Cursor, plus Notion and agentic coding tools like Claude Code

For a power user who likes owning the stack, this is a lot of tool for $8.49: you pick the model, bringing your own keys for OpenAI, Anthropic, Deepgram or Groq, choosing among GPT-5, Claude Haiku 4.5, Llama 4, Grok 4.1, Gemini 3.0 Flash and Ministral, or running Whisper locally and offline, with modes that adapt output to whatever app you're in. Then there's the evidence problem, and it's a real one, because Trustpilot carries only two reviews for the product and one of them says this:

A Trustpilot reviewer reported the app sending all of the context of a Claude Code session into the cloud, with privacy settings disabled.

One report is one report. It lands on a tool whose main selling point is local, private processing, so verify that behavior yourself before you dictate anything sensitive. Reviewers also mention non-literal contextual output where they expected a transcript, and API usage-limit errors during normal use. SOC 2 Type II is claimed for the Enterprise tier, which is quote-only.

19. TurboScribe, unlimited transcription for the price of lunch

Shape: standalone upload-and-export transcriber
Price: Unlimited at $10/month billed yearly ($120/year) or $20/month billed monthly; Teams at $10/user/month
Free: 3 transcripts/day, up to 30 minutes each
Accuracy evidence: Whisper-powered with Cheetah, Dolphin and Whale modes, vendor's own framing; Trustpilot 4.4/5 from 430 reviews praising accuracy and speed

If you have a hard drive of recordings and no engineering time, the math here is hard to argue with. No per-minute meter, files up to 10 hours and 5GB, transcription across 98+ languages, translation into 130+, built-in summaries, speaker recognition, noise reduction, and export to DOCX, PDF, TXT, SRT and VTT. The free tier at three transcripts a day is a genuine plan rather than a countdown.

Can you automate it? Not on the published evidence: no API, Zapier connector or app-ecosystem integration appears on TurboScribe's official marketing and pricing pages. So plan for a person clicking upload, and count that person's time as your cost of adoption. Trustpilot reviewers report unexpected charges and a restrictive refund policy on annual plans, slow or unresponsive support, and processing failures on large or lengthy files despite the stated limits. Encryption of files and transcripts is stated on the site.

20. Verbit, built for the audits

Shape: hybrid AI plus human, aimed at compliance-grade output
Price: Standard from $24/month for 100 hours (50 live plus 50 post-production) and 1 user; Legal Capture, Legal Visor and Campus Complete are quote-only
Free: no free tier and no trial
Accuracy evidence: vendor's own, on Captivate ASR paired with human editing and QA
Wires into: Kaltura, Panopto, Canvas, plus Zoom, Teams, ON24, Vimeo, Brightcove, YouTube, AWS, Box, Dropbox, Google Drive and Blackboard/Anthology

Verbit sells court-admissible transcripts through Legal Capture, case insights through Legal Visor, and ADA Title II accessibility coverage through Campus Complete, and states that it serves 3,000+ leading organizations across legal, media, education, corporate and government. The suite spans live captioning, translation into 50+ languages, audio description and dubbing, with a claimed 5-minute setup. The certification list it publishes is the longest here: ADA, GDPR, FCC, CVAA, WCAG 2.0 AA, ISO 27001, AICPA and SOC 2 Type II. That breadth is the reason to look at all.

Can you try it first? No free tier, no trial. The entry plan buys 100 hours a month for one user, which works out at $0.24 an hour if you use all of it, and every tier above it requires a quote, so your evaluation starts with a sales cycle and a signature rather than a test file. Independent product evidence is close to absent too, and the two Trustpilot reviews that exist concern freelance transcriptionist pay and treatment rather than the product, with allegations of very low pay and non-payment to contractors.

21. Trint

Shape: production suite for journalism and media teams
Price: the self-serve pricing pages do not load, so there is no tier price to read. An API is available on many plans, with Enterprise recommended for high volume
Free: -
Accuracy evidence: transcription across 40+ languages with automatic detection and translation into 70+, vendor's own
Wires into: collaboration and MAM tools, plus the API on many plans

Why is a 2.8 out of 5 in a buyer's guide at all? Because the complaints are about billing, and billing lands on you.

Trustpilot scores Trint at 2.8 out of 5 across 370 reviews. The recurring themes are auto-renewal billing and difficulty cancelling, with one reviewer citing a charge of around £2,000 for a product they were not using.

Now put that next to the price bullet. Every other tool on this page publishes a rate; here, trint.com/pricing and trint.com/plans both return 404 and app.trint.com/plans is a JavaScript-rendered app that renders nothing readable, so there is no figure to quote. Whatever the cause, you go into this one without a published rate in hand, and the loudest complaint about the company is what it charges. The editorial features are a genuine draw: an AI Assistant for summarization, quote extraction and insight identification, captioning, rough-cut creation, and a Story feature reviewers credit with speeding up content repurposing. Security holds up well too: Trint publishes ISO 27001 and Cyber Essentials certification, a choice of EU or U.S. data storage, and a stated policy against training algorithms on customer transcripts. Reviewers still report poor accuracy specifically in French and German despite the multi-language marketing, and slow or unresolved support, including one user left without help for six days.

22. tl;dv: capable feature list, troubling service record

Shape: meeting notetaker for Zoom, Google Meet and Microsoft Teams
Price: Pro €29/month per seat (€18/month annually, €216/year), Business €39/month per seat (€29/month annually, €348/year)
Free: a €0 plan is offered
Accuracy evidence: vendor's own
Wires into: HubSpot, Slack, Zapier, from a claimed 5,000+ tool integrations

On paper it competes: note formats including MEDDIC and Smart AI Topics, automatic CRM logging, drafted follow-up emails, AI sales coaching, trend analysis across meetings, and 30+ languages and dialects, with tl;dv self-reporting 2M+ users and a 4.7/5 G2 rating from 450 reviews. The independent record points the other way.

Trustpilot rates tl;dv 2.2 out of 5 from 9 reviews, 8 of them within the last 12 months. Reviewers describe being completely locked out of paid accounts and recordings, support being ignored for weeks on simple requests, and billing continuing after a cancellation was submitted. One reviewer describes the marketed free offer funnelling into a 7-day trial of the paid Business plan.

Nine reviews is a thin sample and you should weigh it as such. The failures described are account access and money, which are the two things that hurt most when they go wrong mid-quarter. Ranked last for that reason.

Common questions

Which speech-to-text engine actually has the lowest word error rate, and who verified it?

Two candidates, two third-party benchmarks, two different datasets. AssemblyAI's Universal-3.5 Pro topped Coval's independent July 2026 run of 31 models at 3.4% WER, a result AssemblyAI publishes on its own site; on AssemblyAI's own benchmarks the model posts 4.35% normalized WER for pre-recorded audio against 5.338% for OpenAI GPT-4o Transcribe and 6.662% for Deepgram Nova-3. Speechmatics recorded 1.07% WER in Pipecat's August 2026 test of 12 speech-to-text services, the lowest of that group, and that result is readable on Speechmatics' own site. Different datasets and different model sets, so treat these as two credible data points rather than a head-to-head result.

Which of these have a genuinely permanent free plan, and where exactly do the caps bite?

Fathom is uncapped on recordings and transcriptions. Self-hosted Whisper is free under an MIT license. Superwhisper keeps its core features unlimited once the 3,000-word Pro trial is spent. After that the caps arrive fast: Otter at 300 minutes/month, Fireflies at 400 minutes of team storage with a 2-hour meeting limit, Notta at 120 minutes/month and just 3 minutes per conversation, Descript at 60 minutes/month, Rev at 45 AI minutes/month in English on 1 seat, TurboScribe at 3 transcripts/day up to 30 minutes each, Granola at 30 days of history, Wispr Flow at 2,000 words/week on desktop and unlimited on Android, ElevenLabs at 10,000 credits/month which is roughly thirty minutes, and Google at 60 minutes/month on the V1 API. Sonix (30 minutes) and Happy Scribe (10 minutes) are trials. Speechmatics, AssemblyAI, Deepgram and Gladia give credits instead: $100, $50, $200 and €50.

What does transcription really cost per hour of audio once the starter credit runs out?

The fourth column of the first table does the conversion for you, and here are the published rates behind it. AssemblyAI $0.15/hour on Universal-2, $0.21/hour on Universal-3.5 Pro, $0.15/hour streaming and $4.50/hour for the Voice Agent API. Gladia $0.61/hour async and $0.75/hour real-time on Starter, from $0.20 and $0.25 on Growth. Sonix $10/hour pay as you go. Deepgram $0.0048/minute for Nova-3 streaming, which is $0.29 an hour. Whisper $0.006/minute or $0.36 an hour, halving to $0.18 on gpt-4o-mini-transcribe, and $0 self-hosted. Google from $0.016/minute with data logging ($0.024/minute without) down to $0.004/minute above 2,000,000 minutes/month, so $0.96 an hour with logging or $1.44 without, falling to $0.24, with $0.003/minute for dynamic batch and $0.078/minute for medical models. Speechmatics publishes $0.129 per unit with volume discounts, priced per unit rather than per hour, so it takes a conversion of your own before it sits on the same axis. Then check the next tier up: Deepgram's Growth plan starts at $4K+/year, which is the number that actually decides whether the pilot survives.

Which vendors will sign a HIPAA BAA or give me EU data residency without an enterprise contract?

AssemblyAI is the cleanest answer: it lists a HIPAA BAA, PCI-DSS, SOC 2 Type 2 and ISO 27001 as included at no premium, with an EU region priced the same as U.S. Gladia states SOC 2 Type II, GDPR, HIPAA and ISO 27001 on its pricing page from an EU base. Deepgram states HIPAA BAA availability and GDPR readiness. Wispr Flow states HIPAA-readiness with a BAA. Happy Scribe offers EU data residency, Trint offers EU or U.S. storage, and Google Cloud offers regional residency plus customer-managed encryption keys. Sales-gated instead: Otter sells HIPAA only as an Enterprise add-on, Rev puts CJIS/HIPAA on the custom Unlimited tier, Fireflies puts HIPAA/FERPA on the $39/seat Enterprise plan, and ElevenLabs puts the BAA on Enterprise.

Can any of these handle strong accents and overlapping speakers reliably?

Not dependably, no. Read the right-hand column of the accuracy ledger: where a failure mode is named, accents, overlap and noise account for six of the rows. Reviewers report it for Fireflies, Otter, Notta, Sonix, AssemblyAI and Deepgram alike. Two partial counterweights: Gladia's G2 reviewers specifically praise its accuracy on accented speech, and ElevenLabs publishes per-language WER tiers so you can look up where your language actually lands rather than trusting a language count. Wispr Flow's Hindi and Hinglish are called barely functional by reviewers, and Trint draws French and German complaints. Test with your worst audio, not a clean sample.

Can I self-host or run transcription offline instead of paying per minute?

Yes, and at $0 in usage cost it's also the cheapest route on this page. Whisper is MIT-licensed with six model sizes from 39M to 1550M parameters plus an 809M turbo variant that runs about 8x faster than large, and it runs entirely on your hardware. Speechmatics ships as an on-premises virtual appliance as well as Docker and Kubernetes. Google Cloud offers an on-prem deployment option. Superwhisper supports local model processing on the desktop. The trade is engineering time, plus no diarization in classic whisper-1 unless you move to gpt-4o-transcribe-diarize.

Before you sign anything

Run your own worst audio through the top two candidates. Not a clean sample. The call with the bad mic, the two people talking over each other, the accent your model has never met.
Look up the per-language WER tier rather than the language count. A 100+ number tells you what the marketing team wrote; a per-language figure tells you what you'll ship.
Model cost at your real monthly minutes in a single unit, and price the tier above the one you plan to buy. A starter credit gets spent. The commitment tier is the one you renew.
Confirm the BAA, the residency region and the SSO you need are available at the tier you'll actually purchase, in writing, before the pilot ends.
Find the cancellation path before you take the annual discount. On this list, that's where the loudest complaints live.

Which of the two markets are you buying from? Answer that first, because it decides everything else. The accuracy claims on this page run from third-party benchmark numbers to a bare adjective, and the free plans run from genuinely unlimited to three minutes at a time. Do those five checks, know which shelf you're standing at, and the shortlist picks itself.

Explore AI Speech to Text Tools