Transcribing an hour of audio costs about $0.27. That is OpenAI’s published list price for GPT-Transcribe, its batch speech-to-text API, at $0.0045 per minute. Nobody in this category is charging you for transcription. They are charging you for the wrapper around it: the summaries, the action items, the search across a year of calls, the CRM sync, the bot that shows up so you do not have to.
That is worth saying out loud before you compare a single price, because it reframes every decision below. When Sonix charges $10 an hour pay-as-you-go and Happy Scribe charges $12 an hour for extra credits, the underlying compute is a rounding error against the sticker. You are buying software, not minutes.
The second finding is harder to say politely. Not one meeting-notes app appears in any credible independent accuracy benchmark. The two vendor-neutral leaderboards that exist, Artificial Analysis and the Hugging Face Open ASR Leaderboard, benchmark speech recognition engines such as Whisper, ElevenLabs Scribe and MAI-Transcribe. Otter, Fireflies, Fathom, Grain, tl;dv, Granola and Notta appear in neither, and none of them discloses which engine it runs. So every “99% accurate” claim in this category is unfalsifiable by construction. There is no test you could run that would prove it wrong, because there is no disclosed subject to test.
How this guide was made, and what it is not
This is a desk study. Every number below comes from vendor pricing pages, security and privacy policy documents, official help-centre articles, deprecation notices and public benchmark leaderboards, fetched directly on 13 September 2026. We did not run these tools side by side on the same audio and score the transcripts. No hands-on trial happened, so nothing here should be read as one. Where a claim could not be traced to an official page we say so, and there is a full section near the end listing what we could not verify.
Prices render differently by geography, currency, logged-in state and A/B test. The figures below are what each page returned to an unauthenticated US fetch on that date. All figures are as of September 2026.
The $0.27 hour: what transcription actually costs
OpenAI publishes two speech-to-text rates. GPT-Transcribe, the asynchronous batch model, is listed at $0.0045 per minute, which is $0.27 per audio hour. GPT-Live-Transcribe, the streaming model you would need for real-time captions, is listed at $0.017 per minute, or $1.02 per audio hour. Streaming costs roughly four times batch, which is the one genuine cost driver in the stack.
Put that against what the tools charge when you buy transcription by the hour rather than by the seat.
| Route | Cost per audio hour | Source |
|---|---|---|
| OpenAI GPT-Transcribe (batch API) | $0.27 | openai.com/api/pricing |
| OpenAI GPT-Live-Transcribe (streaming API) | $1.02 | openai.com/api/pricing |
| Sonix Pay As You Go, and Sonix overage on every plan | $10.00 | sonix.ai/pricing |
| Happy Scribe extra AI credits ($0.20/min) | $12.00 | happyscribe.com/pricing |
| Rev human transcription ($1.99/min) | $119.40 | rev.com/pricing |
| Self-hosted Whisper (MIT-licensed weights) | $0 licence, plus your own hardware and electricity | github.com/openai/whisper |
The Rev line is not an outlier to be embarrassed about. Rev’s $1.99 per minute buys a human being typing, advertised at “99%+ accurate, delivered in 12 hours or less”. That is a different product, and for a deposition or a broadcast caption it is the right product. It is in the table so you can see the full 440x spread between the cheapest machine route and the human one.
The useful read is the middle of the table. Sonix at $10 an hour and Happy Scribe at $12 an hour are selling machine transcription at roughly 37 to 44 times the batch API rate. That markup is not a scandal; it pays for a web app, speaker labels, translation, storage, support and a business. But it does mean that if all you need is text out of audio, and you have any volume at all, the economics of doing it yourself are not close.
Nobody in this category has ever been independently tested
Here is what the vendors claim on their own pages, and what backs it.
| Tool | Published accuracy claim | Methodology published? | Appears in an independent benchmark? |
|---|---|---|---|
| Sonix | “99% accurate AI transcription” | No | No |
| Happy Scribe | “99% accuracy” — for its human service | No | No |
| Rev | “99%+ accurate” human transcription; “99.7%+ accuracy” legal premium transcript — both human services, not ASR | No | No |
| Otter | No percentage at all | N/A | No |
| Notta | No percentage on its own pricing page | N/A | No |
| Fireflies, Fathom, Grain, tl;dv, Granola | No percentage found on official pages | N/A | No |
Otter deserves credit for the one honest position here. It publishes no number, and its help centre says only that its recognition is “generally quite accurate”, adding that accuracy “can vary based on factors such as background noise, speaker accents, and the complexity of the conversation.” That is what every vendor in this category could truthfully say. Most of them say 99% instead.
Rev’s “independent benchmarking study”
Rev publishes something called The 2024 State of ASR Report, which describes itself as an independent benchmarking study and says it evaluated nine leading ASR providers. The claimed margins are specific: Rev beats Google by 60.5%, Otter by 24.1%, Amazon by 5.6% and Microsoft by 5.3%, and is 47% better than competitors in challenging recording environments.
Read the funding line. There is not one. The report does not state who conducted it or who paid for it. It is hosted and published by Rev. Rev wins. And it is a 2024 report, which in a field where engine quality moves every few months makes it two generations stale. We are not saying the numbers are wrong. We are saying that a study with an undisclosed funder, published by the party it ranks first, is marketing with a methodology section.
This pattern is not unique to Rev. Almost every “2026 independent transcription accuracy test” that surfaces in search is published by a transcription vendor. LecSync, VexaScribe, MirrorCaption, MeetMemo, VidNotes, Gladia, Sonix, Notta and tl;dv all run competitor comparisons on their own blogs. If a comparison ranks the company hosting it first, you have a brochure.
The two benchmarks that are real, and why your app is not in them
There are genuine vendor-neutral leaderboards. The Artificial Analysis Speech-to-Text leaderboard covers 58 models using its Word Error Rate Index over roughly 8 hours of audio, weighted across AA-AgentTalk at 50%, VoxPopuli-Cleaned-AA at 25% and Earnings22-Cleaned-AA at 25%, deliberately spanning accents, domain-specific language and difficult acoustic conditions.
| Model | AA-WER v2 (non-streaming, lower is better) |
|---|---|
| Fun-Realtime-ASR-preview | 1.7% |
| MAI-Transcribe-2 | 2.0% |
| ElevenLabs Scribe v2 | 2.2% |
| MAI-Transcribe-1.5 | 2.4% |
| Smallest AI Pulse Pro | 2.4% |
The second is the Hugging Face Open ASR Leaderboard, which as of its November 2025 write-up compared 60-plus open and closed models from 18 organisations across 11 datasets, community-maintained through a public GitHub repository with named contributors. It reports that OpenAI’s Whisper Large v3 performs best among open models on long-form audio. The specific word error rates for individual Whisper variants were not in the material we fetched, so we are not quoting a number for Whisper.
Now look at what is in those two lists: engines. Whisper, Scribe, Canary, Chirp, MAI-Transcribe. Not one meeting-notes app is there. That is not an oversight by the benchmark maintainers. Otter, Fireflies, Fathom, Grain, tl;dv, Granola and Notta are applications built on top of engines, and none of them tells you which engine. You cannot benchmark a black box against a public dataset when the vendor will not say what is in the box, and you cannot disprove a claim that was never made falsifiable.
The practical instruction: ignore accuracy claims entirely when choosing in this category. Choose on privacy posture, integration, price mechanism and whether a bot in the room is acceptable to your counterparties. Those are things you can actually verify.
Which tools train on your meetings
This is the most decision-relevant section in the guide, and the one where the answers genuinely differ. Three of the most popular tools in this category train on your meeting content. Two of those give you an opt-out. One of them we could not find a documented opt-out for at all.
| Tool | Trains on your content? | Control you get | Certifications stated on an official page |
|---|---|---|---|
| Otter | Yes — policy describes “training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)”, with legal basis given as consent or legitimate interests | No documented opt-out found in the policy — could not verify | HIPAA compliance announced 10 July 2025 |
| Fathom | Yes, de-identified, to train in-house models; third parties explicitly barred | Opt-out in account settings | SOC 2 Type II / HIPAA / GDPR appear only in the page meta description, not the policy body — treat as unverified |
| Granola | Yes, on anonymised data, for product improvement; third parties barred | Opt-out in Settings | SOC 2 Type 2 via independent audit; GDPR with DPA on request. HIPAA not mentioned |
| Descript | Only if you choose to share | Opt-in, disabled by default | SOC 2 Type II; aligns with GDPR, CCPA, Privacy by Design. No HIPAA |
| Fireflies | No, and vendors contractually barred from training on it | N/A — blanket prohibition | Policy references SOC2, HIPAA, GDPR and the EU/Swiss/UK Data Privacy Framework |
| Rev | No — does not sell customer data or allow it to train external LLMs | Blanket | SOC 2 Type II; HIPAA (enterprise tier); GDPR; CJIS (unlimited plan); PCI |
| Sonix | No — “None of your data processed through Sonix is used for training purposes” | Blanket | SOC 2 Type II; HIPAA via Medical Sonix with BAAs; GDPR DPA and SCC |
| Zoom | No — does not use customer audio, video, chat or screen share to train its own or third-party models | Blanket, per its September 2023 commitment | — |
| Microsoft | No — prompts, responses and Graph data are not used to train foundation LLMs | Blanket | — |
| Not without permission — no training and no human review outside your domain without prior permission or instruction | Permission-gated | — | |
| Happy Scribe | Could not verify — training is not addressed anywhere on its security page | — | SOC 2 Type 2; GDPR; EU data centre described as ISO 27001 compliant. HIPAA not mentioned |
| Grain, tl;dv, Notta | Could not verify from official security pages in this pass | — | — |
Three things in that table are worth pulling out.
The tools that train on you are the popular ones. Otter, Fathom and Granola are three of the best-known products in the category, and all three use your meeting content to improve their models. The word “de-identified” or “anonymised” is doing real work in each policy, and it is a meaningful protection, but it is not the same as “we do not use your meetings”.
Default settings matter more than policy language. Descript is the only one of the four that is off unless you turn it on. Fathom and Granola are on unless you turn them off, which means the modal user is training the model. If you deploy either across a team, the opt-out is a per-account setting somebody has to actually go and set.
Otter is the gap. Its privacy policy describes training on de-identified audio and on transcriptions that may contain personal information, and we could not find a documented toggle to switch that off. That is not proof no toggle exists. It is a statement that we looked at the policy and did not find one described, which for a product with HIPAA compliance announced in July 2025 is a strange thing to leave undocumented.
The flat-no group is genuinely reassuring, and it splits along an interesting line. Fireflies, Rev and Sonix are independent vendors that chose a blanket prohibition as a differentiator. Zoom, Microsoft and Google give the same answer because they sell you a platform and your meeting content is not their product. If your compliance team’s only question is “does the vendor train on our data”, that group is your shortlist.
Retention is almost universally undefined
Ask a simpler question than training: how long does this company keep your meetings? Almost nobody will tell you.
| Tool | Stated retention commitment |
|---|---|
| Fireflies | Account data deleted within 30 days of account closure. The only specific number found in this research pass |
| Otter | No duration given. Policy says data is kept “as long as necessary to fulfill the purposes set out in this Policy” |
| Fathom | Worded near-identically to Otter. No duration given. Storage described as “unlimited”; exact retention could not be verified |
| Granola | No retention period stated on its security page. Notes stored in a US-hosted AWS Virtual Private Cloud, encrypted at rest and in transit, backed up daily |
| Sonix | No deletion window stated, but promises continued access: “Your files are always accessible, even after your subscription ends” |
| Everyone else | Not stated on the pages fetched — could not verify |
“As long as necessary to fulfill the purposes set out in this Policy” is the standard formula, and it is not a commitment. It means the company decides, it can change its mind, and you have no contractual number to point at. For most people that is tolerable. For anyone whose meetings contain client confidences, medical detail, HR matters or anything that could be discovered in litigation, the absence of a retention clock is the whole problem: you cannot promise anyone that a recording goes away if the vendor has not promised you.
Fireflies being the only vendor to commit to a number is a low bar cleared by exactly one company. Sonix’s position is the interesting inverse: it is not promising deletion, it is promising the opposite, that your files stay available after you stop paying. That is a genuine feature if you are archiving a back catalogue, and a genuine liability if you wanted the material gone.
The free tiers cannot be compared, because they meter three different things
Every roundup in this category lines up free plans in a column and implies they are alternatives to each other. They are not. The free tiers limit three incompatible quantities: minutes you may transcribe, minutes you may keep, and how far back you may look.
| Tool | Free tier | What is actually capped |
|---|---|---|
| Otter Basic | 300 monthly transcription minutes; 25 most recent conversations retained; 30 minutes max per conversation | Minutes and history and per-meeting length |
| Notta Free | 120 transcription minutes/month; up to 3 minutes per conversation; 50 file uploads/month; 10 AI summaries/month | Minutes, and a 3-minute per-conversation ceiling that rules out real meetings |
| Rev Free | 45 AI minutes per user per month; English only; 1 user | Minutes |
| Descript Free | 60 minutes of media per month; 100 one-time AI credits | Minutes |
| Happy Scribe Free | 10-minute trial; 45-minute maximum recording duration; limited history | A one-off trial, not a recurring allowance |
| Sonix | 30-minute free trial, no credit card | A one-off trial |
| Fireflies Free | Transcription advertised as unlimited, but 400 minutes of storage for the whole team | Storage, not transcription. The cap is on keeping, not making |
| Fathom Free | “Unlimited recordings + transcriptions”, free forever | Nothing obvious on volume; features are the gate |
| Granola Basic | “Unlimited meeting notes for free”, but you must upgrade to view or work with notes older than 30 days | A 30-day history window, not a minute cap |
| Grain Free | 20 meetings | Meeting count |
| Zoom Basic | Meeting summaries for 3 hosted meetings per month | Meeting count |
| tl;dv | Help centre describes “a free plan, with unlimited recordings and transcripts” with limits on other features. Exact caps could not be verified | — |
Three mechanisms, and they fail you in different ways. A minute cap bites when you have a lot of meetings. A storage cap bites when you have a few long ones. A history window bites when you go back to look for something, which is the entire reason most people install these tools in the first place.
Granola and Fireflies are the two worth understanding before you build a habit on them. Granola will happily transcribe everything forever and then, on day 31, put a paywall between you and the note you are trying to find. Fireflies advertises unlimited transcription on the free plan and caps team storage at 400 minutes, so the constraint arrives after the work is done rather than before.
Notta’s three-minute per-conversation ceiling on the free plan is not a meeting tool at all; it is a demo. Otter’s free plan, at 300 minutes a month with 30-minute meetings and only the 25 most recent conversations kept, is the most usable of the genuinely metered ones, and it also trains on your content.
What the paid plans cost
Prices below are as published on each vendor’s own pricing page on 13 September 2026. Read the caveats under the table before you use any of them in a spreadsheet.
| Tool | Entry paid tier | Next tier up | Notes |
|---|---|---|---|
| Otter | Pro: $16.99/user/month monthly, $8.33/user/month annual (page also showed promo figures of $8.49 and $4.17) | Business: $30/user/month monthly, $19.99/user/month annual | Pro gives 1,200 in-app recording minutes, 90 min/meeting; Business unlimited meetings, 4 hrs/meeting |
| Fireflies | Pro: $18 monthly, $10/seat/month billed annually | Business: $29 monthly, $19/seat/month annual; Enterprise $39/seat/month annual | Pro gives 8,000 minutes of storage per seat; Business unlimited |
| Fathom | Premium: $20 monthly, $16/month annual | Team: $19/user monthly, $15/user annual (min 2); Business: $34/user monthly, $25/user annual (min 2) | All plans carry a 90-day guarantee |
| Granola | Business: $14/user/month | Enterprise: $35/user/month | Annual pricing not shown on the page — could not verify |
| Grain | Starter: $19/seat/month billed annually | Business and Enterprise: prices not displayed | Enterprise routes to “Book a demo” |
| Notta | Pro: $8.17/month annual (total $97.99 billed annually) | Business: $16.67/month annual (total $199.99); Enterprise custom from 51 seats | Month-to-month rates not stated on the page — could not verify |
| Descript | Hobbyist: $16 and $24 per person/month | Creator: $24 and $35; Business: $50 and $65 | Which column is annual is genuinely unclear — see below |
| Rev | Essentials: $25.49/month or $29.99/month annual, 5,000 AI minutes, English and Spanish, 3 users | Pro: $47.99/month or $59.99/month annual, 10,000 minutes, 37+ languages, 5 users | The column labelling on Rev’s page is odd — see below |
| Sonix | Core: $25/mo or $275/yr (5 hrs/mo, 60 hrs/yr) | Advanced: $50/mo or $550/yr; Pro: $80/mo or $880/yr | Pay As You Go at $10/hr with no subscription; overage on all plans at $10/hr |
| Happy Scribe | Basic: $8.50/month annual, $17/month monthly, 120 AI minutes/month | Pro: $19 annual / $29 monthly, 600 min; Business: $59 annual / $89 monthly, 6,000 min | Extra AI credits $0.20/min on all paid tiers |
| Microsoft Teams Premium | $10.00 user/month, paid yearly | — | Includes intelligent meeting recap and intelligent call recap |
| Zoom ZoomMate | Price not published to unauthenticated visitors | — | Page rendered a bare “/user/month” with the numeral absent from the markup |
| Google Meet “Take notes for me” | Requires an eligible Workspace edition or Google AI plan | — | Google’s help page does not enumerate which editions — could not verify |
| tl;dv | Could not verify | — | tldv.io/pricing returns 404; the live pricing page is disallowed by tl;dv’s own robots.txt |
Two pricing pages that contradict themselves
Descript. Two separate fetches both returned the lower figure labelled “monthly” and the higher labelled “annual”. That contradicts a line on the same page promising savings of “up to 35% over monthly plan”. The commercially sensible reading is that the lower number is the annual-billed monthly rate and the higher is month-to-month, which is the opposite of the labels. We could not verify which label belongs to which column, so do not budget a Descript annual-versus-monthly split off this table without checking the page yourself.
Rev. Same shape of problem. On Rev’s pricing page the lower figure is labelled monthly and the higher is labelled annual. We reproduce both, and we could not verify which is the true annual-billed rate.
These are not typos we are being pedantic about. They are the numbers you would put in a procurement request, and both vendors publish them in an order that does not match their own savings claims.
Languages and speaker labels
Language coverage is the one spec in this category with a wide, verifiable spread, from six languages to 150-plus. Speaker diarization, the feature that decides whether your transcript reads as a conversation or a wall of text, is much less consistently documented.
| Tool | Languages (per official pages) | Speaker identification |
|---|---|---|
| Happy Scribe | AI transcription 150+; AI translation 80+; human proofreading 65+ | Automatic speaker detection on all tiers |
| Fireflies | 100+ | Not stated on the pricing page — could not verify |
| Notta | 58 (monolingual transcription, across all transcription features) | On all paid tiers; not listed for Free |
| Sonix | 54+ for transcription and translation; translation into 55+ | Speaker labels, timestamps and speaker voiceprint |
| Fathom | 38, enumerated in its help centre. AI summaries auto-translate for only 6 (Spanish, Portuguese, German, French, Italian, Dutch) on Premium/Team | Not stated on the pages fetched — could not verify |
| Rev | English only on Free; English and Spanish on Essentials; “37+ languages, including Spanglish” on Pro and Unlimited | — |
| Descript | 25 | Detect 8+ speakers on Free, Hobbyist, Creator and Business |
| Google Meet “Take notes for me” | 8: English, French, German, Italian, Japanese, Korean, Portuguese, Spanish | — |
| Otter | 6: English, Spanish, French, German, Japanese, Chinese | Speaker identification by name on Pro, Business and Enterprise |
| Granola | “Multi-language support” claimed, no list published — could not verify | Could not verify |
| Grain, tl;dv | Not stated on the pages fetched — could not verify | Could not verify |
Two things here deserve more weight than they usually get. Otter’s six languages is a genuine constraint for any team that is not working in English, Spanish, French, German, Japanese or Chinese, and it sits oddly against a product now positioning itself as enterprise conversational infrastructure.
Google Meet’s limit is stricter still, and not just in count. Its documentation states plainly that the feature “supports one language at a time” and that multiple languages spoken in the same meeting are not currently supported. If your calls code-switch, which is normal in a lot of the world, that is a hard failure rather than a degradation. Google also recommends the feature for meetings of 15 minutes to a maximum of 8 hours, warning that shorter meetings may produce incomplete summaries.
Diarization being undocumented for Fireflies, Fathom, Grain and Granola is a small scandal of its own. These are meeting tools. Whether they can tell two speakers apart is not a minor spec, and four of them do not state it on an official page.
The Whisper inversion: the API is now the volatile option
The largest single disruption in this category in 2026 has nothing to do with meeting apps. On 26 August 2026, OpenAI notified developers that whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize are deprecated and will be removed from the API on 26 February 2027. The replacements are gpt-transcribe and gpt-live-transcribe.
An enormous amount of tooling is built on the Whisper API endpoint. If you have a script, an integration or a vendor that calls whisper-1, you have until late February 2027 to migrate.
Now the part that inverts the usual assumption. The open Whisper weights are unaffected. They are released under the MIT licence, they are freely downloadable, and no company can deprecate them out from under you. So as of September 2026, in this one narrow case, self-hosting is the stable option and the hosted API is the volatile one. That is the reverse of how build-versus-buy normally runs, and it is worth saying because most teams reach for the API precisely to avoid maintenance risk.
| Model size | Parameters | Required VRAM | Relative speed |
|---|---|---|---|
| tiny | 39 M | ~1 GB | ~10x |
| base | 74 M | ~1 GB | ~7x |
| small | 244 M | ~2 GB | ~4x |
| medium | 769 M | ~5 GB | ~2x |
| large | 1550 M | ~10 GB | 1x |
| turbo | 809 M | ~6 GB | ~8x |
The headline requirement of roughly 10GB of VRAM for the large model is what puts most people off, and it is out of date in practice. The faster-whisper project publishes benchmark runs showing what an optimised implementation does on ordinary hardware.
| Setup | Implementation | Time | Memory |
|---|---|---|---|
| GPU, large-v2, RTX 3070 Ti 8GB, CUDA 12.4 | openai/whisper fp16 | 2m23s | 4,708 MB |
| GPU, large-v2, RTX 3070 Ti 8GB, CUDA 12.4 | faster-whisper fp16 | 1m03s | 4,525 MB |
| GPU, large-v2, RTX 3070 Ti 8GB, CUDA 12.4 | faster-whisper int8 | 59s | 2,926 MB |
| CPU only, small model, Intel Core i7-12700K, 8 threads | openai/whisper fp32 | 6m58s | 2,335 MB |
| CPU only, small model, Intel Core i7-12700K, 8 threads | faster-whisper fp32 | 2m37s | 2,257 MB |
| CPU only, small model, Intel Core i7-12700K, 8 threads | faster-whisper int8 | 1m42s | 1,477 MB |
The faster-whisper repository claims it runs “up to 4 times faster than openai/whisper for the same accuracy while using less memory”, and its own published runs are consistent with that. The practical read: an 8GB consumer GPU runs large-v2 in under 3GB of VRAM with int8 quantisation, and a modern desktop CPU with no GPU at all runs the small model comfortably. For occasional use, no hardware purchase is required.
One honest caveat on that table. The length of the audio clip used for those benchmark runs was not in the material we fetched, so the absolute times mean nothing on their own. The ratios between rows are the useful part.
Where self-hosting wins, and where it does not
Self-hosting genuinely wins in three situations. First, audio that legally cannot leave the building: health, legal, criminal justice, or EU employee recordings restricted by a works council. No DPA negotiation, no sub-processor list, no vendor security review. Second, large back-catalogue batch jobs. At Sonix’s $10 an hour, 500 hours of archive costs $5,000; the same job on the batch API costs $135, and on hardware you already own it costs electricity. Third, anything that has to keep working after February 2027.
Where it loses, and this matters more than the cost argument: raw speech recognition is not a meeting-notes product. You get a transcript. You do not get diarised speakers with names attached, summaries, action items, CRM sync, search across a year of calls, or a calendar integration that knows which meeting this was. At $0.27 an hour, the build-versus-buy decision in this category is essentially never about transcription cost. It is about control, and about the software wrapped around the transcript.
Recording consent: the rule is not one rule
This is not legal advice. We are not lawyers, this section describes the shape of a problem rather than giving you a rule, and we did not verify any non-US law in this research pass. If you record calls at any scale, ask an actual lawyer in the relevant jurisdictions.
The US federal baseline is one-party consent. The Digital Media Law Project’s guide states that federal law permits recording telephone calls and in-person conversations with the consent of at least one of the parties. That is the rule most people have heard, and it is where most people stop.
State law is where it breaks down. The same guide names eleven all-party-consent states, listing them as California, Connecticut, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania and Washington. Count those names: there are ten of them for eleven states. The discrepancy is in the source, not in our transcription of it, and it is a good illustration of why you should not lift a state list off a web page and treat it as settled. Verify state by state.
The exceptions are messy even inside that list. Per the same source, Illinois’s statute was ruled unconstitutional in 2014. Hawaii requires two-party consent only for recordings in private places. Massachusetts prohibits secret recordings rather than demanding explicit consent from everyone. These are not edge cases; they are three of the states people most often cite.
The trap specific to meeting tools is that a video call routinely has participants in several states and often several countries at once, so the strictest jurisdiction on the call effectively governs it. A host in a one-party state such as Texas, recording a colleague in California, is the standard failure mode, and it happens every day inside companies that believe they have checked the rules.
Outside the US this is not wiretapping law at all. In the EU and UK it is data protection law: a GDPR lawful basis, employee-monitoring rules, works council consultation. Different instrument, different analysis, different remedies. We did not verify country-by-country international rules in this pass, so we are not going to generalise about them.
The defensible advice is procedural rather than legal. Get affirmative consent on the record at the top of every recorded call. Leave the tool’s join announcement and recording indicator switched on rather than suppressing them. Let anyone object and honour it. None of that makes you compliant by itself, but all of it is evidence you behaved reasonably, and it costs nothing.
The inversion: on-device recorders look private and make consent invisible
The tools in this category capture audio in three different ways, and the privacy intuition most people have about them is backwards.
| Approach | Tools | Wins on | Loses on |
|---|---|---|---|
| Bot-joiner — joins the call as a participant | Otter, Fireflies, Fathom, Grain, tl;dv | Consent visibility (a named participant in the roster is the clearest possible notice); attending meetings you skip; one tool across Zoom, Meet and Teams | Social friction with clients; enterprise policies that block external app participants outright; captures meetings only, never in-person or phone conversations |
| On-device recorder — captures your computer’s audio, no bot | Granola | Nothing to whitelist; works on any audio your machine can hear, including in-person; survives IT bans on meeting bots | Consent becomes invisible; your machine must be present and awake; capture quality depends on your local audio path |
| Platform-native — built into the meeting platform | Zoom ZoomMate, Teams intelligent recap, Google Meet “Take notes for me” | No third-party data processor at all; no extra vendor security review; no bot to explain | Single-platform lock-in; Google Meet’s 8-language, one-language-at-a-time ceiling; Teams recap depends on transcription being enabled in policy; Zoom’s new credit metering |
Here is the inversion worth naming. A bot in the participant list is the thing everybody finds slightly rude, and it is also the single best consent artefact in the category. “Otter.ai Notetaker has joined the meeting” is unambiguous notice to every person on the call, visible in the roster, timestamped in the log. In an all-party-consent state that is close to the safest posture available.
Granola’s pitch is that it “uses your computer audio, so doesn’t invite a bot” and “doesn’t join your meeting. Instead, it transcribes in the background.” That reads as the privacy-respecting option, and in one sense it is: no third party sits in the room. But nothing in the roster tells the other side that a transcript is being made. The people on the call have no way to know, no way to object, and no record that they were told. That is exactly the risk profile the all-party-consent states exist to address, and it is a genuine ethical and legal downside, not a quirk.
None of which makes on-device recording wrong. It makes the announcement your job instead of the software’s. If you use a tool that does not announce itself, you have to.
Teams has a dependency worth knowing before you buy: if the transcription setting is toggled off in your meeting policies, intelligent meeting recap is simply unavailable to users under that policy. The feature does not degrade, it disappears.
What changed in 2025 and 2026
Beyond the Whisper deprecation, four shifts are worth knowing before you sign anything.
Zoom’s AI went from free to metered. In September 2023 Zoom announced that AI Companion was “now included at no additional cost for customers with the paid services in their Zoom user accounts”. As of 2026 the assistant page describes a free Basic tier limited to meeting summaries for three hosted meetings per month, and a paid ZoomMate tier including 2,200 AI credits recurring monthly. Zoom support confirms the metering: metered AI features require both a ZoomMate licence and available credits, credits pool at account level, monthly credits do not roll over, credit packs expire one year from purchase, and per-action cost varies by prompt, complexity, model used and data volume. Zoom’s own knowledge base declines to state allocations, pointing you to the pricing page or an account executive. This is the single most important pricing shift in the category, and it is a move from a flat inclusion to a consumption model.
Microsoft loosened intelligent recap licensing. As of a documentation revision dated 10 April 2026, a Microsoft Copilot licence now satisfies the requirement alongside Teams Premium for 1:1 Teams calls, and for PSTN a Teams Phone licence plus either Teams Premium or Copilot. If your organisation already pays for Copilot, check before you buy Teams Premium for recap alone. Teams Premium itself lists at $10.00 per user per month paid yearly.
Two vendors made pricing harder to see. Grain now publishes only Free at $0 and Starter at $19 per seat per month billed annually; Business and Enterprise prices are not shown, with Enterprise routing to a demo booking. tl;dv has gone further: tldv.io/pricing returns a 404 and the live pricing page is disallowed by tl;dv’s own robots.txt, so it cannot be fetched or indexed. The only official statement we could obtain was a help-centre article last updated 3 April 2025 describing a free plan with unlimited recordings and transcripts, plus Pro and Business plans with unlimited AI features and more advanced capabilities. We are not quoting a tl;dv price, because there is no official one to quote.
Everything else is growing, not dying. Otter announced HIPAA compliance on 10 July 2025, a $100M ARR milestone on 22 December 2025, a Japanese-market launch on 4 November 2025, Live Assist on 21 July 2026, and an April 2026 repositioning away from “AI notetaker” toward enterprise conversational knowledge. Google extended “Take notes for me” to in-person meetings in August 2026. Fathom moved its primary domain from fathom.video to fathom.ai, with the old pricing URL now 302-redirecting. And a Bloomberg headline dated 25 March 2026 reports Granola hitting a $1.5 billion valuation on $125 million in funding, although we could only verify the headline; the article body is paywalled.
No shutdowns were confirmed among Otter, Rev, Descript, Fireflies, Grain, Fathom, tl;dv, Notta, Sonix, Happy Scribe, Granola, Zoom, Teams or Google Meet. Every product here still exists as sold. The only thing genuinely being retired is OpenAI’s Whisper API endpoint.
What we could not verify
This section is deliberate, and on this topic it is unusually long. These are things a buyer’s guide is expected to tell you that we could not source to an official page. We would rather list them than fill them in from memory.
- tl;dv’s entire pricing table. The public pricing URL 404s and the live page is blocked by tl;dv’s own robots.txt. Needs a manual browser check.
- ZoomMate’s per-user price and per-plan AI credit allocations. Zoom does not publish them to unauthenticated visitors, and every figure circulating online traces to third-party blogs.
- Grain’s Business tier price, and Grain’s language support, retention and diarization, none of which appear on its pricing page.
- Descript’s annual-versus-monthly column mapping. The page contradicts its own savings claim. Needs a screenshot to settle.
- Rev’s annual-versus-monthly column mapping, for the same reason.
- Notta’s month-to-month prices. Only annual totals are stated; the monthly rate is implied by a 40% discount figure and never given.
- Diarization for Fireflies, Fathom, Grain and Granola. Not confirmed on any official page for any of the four.
- Which Google Workspace editions include “Take notes for me”. Google’s help page requires “an eligible” edition and does not enumerate them.
- Fathom’s SOC 2 Type II, HIPAA and GDPR claims. These appear only in the privacy page’s meta description, not in the policy body.
- Happy Scribe’s AI-training stance. Not addressed anywhere on its security page.
- Whether Otter offers any opt-out from model training. The policy describes the training; we found no documented toggle.
- Fireflies’ specific security certificates. The privacy policy references SOC2, HIPAA, GDPR and the Data Privacy Framework, but the Trust Center page returned only the line “You own your data, and we never use it for training”.
- Granola’s language count and retention period. Multi-language support is claimed with no list; no retention period is stated.
- Specific word error rates for Whisper variants on the Hugging Face Open ASR Leaderboard. The write-up names Whisper Large v3 as the best open model on long-form audio without numbers we could quote.
- The audio clip length behind the faster-whisper benchmark runs. Treat those times as ratios only.
- Any 2025 or 2026 acquisition of Descript, Grain, tl;dv or Rev. Searching surfaced only aggregator and profile pages, no primary announcement. We are not asserting one.
- International recording-consent rules. Deliberately out of scope for this pass. Needs its own research or a lawyer.
- Hands-on accuracy of any tool here. We did not test them. Nobody credible has.
The verdict
Five things to take away, in the order they should change what you do.
Stop shopping on accuracy. There is no independent evidence for any app-level accuracy claim in this category and there cannot be while the vendors keep their engines undisclosed. Rev’s benchmark is published by Rev, funded by nobody it will name, and ranks Rev first. Treat 99% as a font choice.
Shop on training posture and defaults instead. If your meetings contain anything confidential, the flat-no group is your shortlist: Fireflies, Rev and Sonix among the independents, and Zoom, Microsoft and Google among the platforms. If you want Fathom or Granola, deploy the opt-out as part of onboarding rather than hoping people find it. If Otter is the pick, go in knowing it trains on de-identified audio and transcriptions, and that we could not find a documented way to stop it.
Check what you already own before buying anything. If you pay for Microsoft Copilot, that licence now satisfies intelligent recap and you may not need Teams Premium at $10 per user per month for recap alone. If you are on Workspace, “Take notes for me” may already be included, subject to an edition list Google does not publish. Single-platform companies routinely buy a third tool for a feature already in the bundle.
Match the capture method to your consent exposure, not to your taste. Regulated industries and teams with people in all-party-consent states should prefer a bot, precisely because the bot is visible in the roster and creates an audit trail. Consultants, anyone with in-person meetings, and anyone whose clients block external participants should prefer on-device, and should then announce the recording themselves, because the software will not.
If your need is volume transcription rather than meeting software, do it yourself. At $0.27 per audio hour on the batch API, or free on MIT-licensed weights that run in under 3GB of VRAM on a mid-range gaming GPU, paying $10 to $12 an hour for machine transcription only makes sense when you are actually using the software around it. And after 26 August 2026, the local weights are the durable choice: the API endpoints have a removal date of 26 February 2027 and the weights do not have one at all.
All figures here are as of September 2026 and were taken from vendor pages that change without notice. Two of them contradict themselves, two more decline to publish prices at all, and one is blocked from being read. Re-check anything you are about to spend money on. For more comparisons built the same way, see our full guide index, our screen recording tools comparison, which covers a lot of the same capture and storage tradeoffs, and our email client roundup.
Leave a Reply