Local vs cloud meeting transcription

Where does your audio actually go?

With a cloud notetaker, your meeting audio is streamed or uploaded to the vendor's infrastructure — commonly Amazon or Google cloud in the United States — where it's transcribed and the transcript is stored in your account. Encryption in transit and at rest (typically TLS and AES-256) is standard, and reputable vendors hold SOC 2 and similar certifications. The point isn't that this is reckless; it's that the record of your meeting lives on hardware you don't control, under a policy that can change.

With local, on-device transcription, the recording is captured and kept on your own machine. Depending on the setup, the speech-to-text step either runs entirely on your computer, or the audio is sent to a transcription provider you choose, under your own account, and the result comes back to your device. Either way there's no vendor-held library of your past meetings by default.

The three questions that actually matter

Marketing pages rarely put these side by side, so here they are plainly.

Question	Typical cloud notetaker	Local / self-hosted (Earshot)
Where do transcripts live?	Vendor's cloud, usually US	Your PC, in a folder you control
Is the audio kept?	Often stored; sometimes deleted post-transcription	Stays on your disk; deleting removes it
Data residency choice?	Frequently US-only, or enterprise-only	Wherever your computer is
Trains the vendor's own models?	Several do, on de-identified data, by default (opt-out)	No — providers you pick are bound by no-training API terms
Needs a vendor account?	Yes	No, when self-hosting

The AI-training question, without the hand-waving

This is the row people get wrong in both directions. The big language-model providers behind these products — the OpenAI or Anthropic APIs doing the summarising — are generally barred by their business terms from training on data sent through the API. That part is reassuring and true for most tools.

The nuance is the notetaker's own models. Several popular cloud services state that they improve their in-house speech or summarisation models on de-identified customer data by default, with an opt-out buried in settings. "De-identified" is doing a lot of work in that sentence, and "by default" means it's happening unless you found the toggle. If that matters to you, read the specific vendor's current privacy page — and check whether the opt-out is on you or on by default.

Local self-hosting sidesteps the question entirely: your audio only ever reaches the transcription and AI provider whose key you supplied, under that provider's no-training terms, and nothing is pooled into a notetaker company's training set.

When cloud is the right choice

Local isn't automatically better for everyone. Cloud notetakers earn their keep with zero setup, cross-device sync, shared team workspaces, and CRM integrations that a local-first tool has to work harder to match. If you're a sales team living in a shared workspace, or you switch between five devices a day, that convenience is real and worth paying for. The trade you're making is control and data residency for smoothness — a fair trade, as long as you're making it on purpose.

When local is the right choice

Local, on-device transcription is the better fit when the content of the meeting is the sensitive part: legal and client-confidential calls, health or financial discussions, HR conversations, product and M&A talks, or any organisation with a data-residency obligation (GDPR, or a policy that says "customer data stays in-region"). It's also the natural choice for people who simply prefer that their own conversations not accumulate in a vendor's account. Recordings are personal data; keeping them on hardware you control is the easiest version of that story to defend — see our guide to meeting recording and consent laws.

How Earshot does it

Earshot records both sides of the conversation locally on your Windows PC, with no bot in the call, and stores the recordings and notes on your machine. There are two ways to run the AI behind it:

• Self-host, free and open source. Point Earshot at your own transcription and AI — a local Whisper server (nothing leaves your computer), Groq's free tier, Deepgram, OpenAI, Anthropic, or a fully local model via Ollama. You pay providers directly under your own keys. Here's the self-hosting guide.
• Earshot Plus, US$9/month. We run the transcription and AI so you don't manage keys. Audio passes through our proxy, is processed, and is discarded — never stored, never used to train models. A 7-day free trial is included, and you can move back to self-hosting whenever you like.