Stored Locally Is Not the Same as Processed On-Device

Two different questions

Stored locally answers where the finished text rests. Processed on-device answers whether your audio ever travels at all. Both matter, but they are not the same promise. When an app says your dictations are private because they are stored on your device, ask the second question: where was the audio turned into text?

This is the same fault line running through the meeting-notetaker stories: the architecture decides what is possible, and the privacy policy only describes what is promised. A policy can say audio is not stored after processing. Only the architecture can say the audio never left.

What Fireflies actually built

Talk is dictation inside the Fireflies desktop app for Mac and Windows, with iOS and Android planned. You hold a key, speak, and release, and clean text appears wherever your cursor is: email, Slack, docs, prompts. It supports more than 90 languages, formats a dictated email like an email and a Slack message like chat, and turns a spoken teammate name into an @mention. It is free and unlimited on every Fireflies plan, which undercuts standalone dictation apps that charge a monthly subscription for the same job.

On privacy, the company says completed dictations are stored locally and cannot be viewed by Fireflies or a workspace administrator, are not stored on its servers after processing, and are not used to train AI models. Its supplemental terms add detail worth reading: audio, transcribed text, and optional Screen Context are sent to Fireflies and its listed service providers for processing. Screen Context can be switched off. Fireflies also advises against dictating while viewing passwords, payment details, or protected health information.

Why the distinction matters

Audio in transit is audio that can be intercepted, logged, or mishandled by a third party, no matter how careful the storage policy is after the fact. Fireflies says dictations are not stored on its servers after processing and are not used to train models, and those are meaningful commitments. But the processing step still happens on someone else's infrastructure, under someone else's controls.

There is also a practical angle. A dictation feature that needs a connection stops working when the connection does. On-device processing keeps working in airplane mode, on hotel wifi, or anywhere you would rather not send your voice in the first place.

The on-device alternative

Echo Chamber was built on the second definition. Transcription runs on on-device models, it works in airplane mode, and nothing is sent to an external AI service for transcription. Recordings and transcripts stay on your iPhone, iPad, or Mac. There is no account, no ads, and no third-party analytics.

Speech recognition runs on the device, and the AI polish only ever touches punctuation, capitalization, and paragraphs, never your words. You can test the claim the same way you would test any privacy promise: turn on airplane mode after setup and dictate. If the words still appear, the processing is local.

It is free for 60 minutes of recording per month. Pro is $2.99/month, $29.99/year, or $79.99 lifetime.

Questions to ask any dictation app

• WHERE IS MY AUDIO TURNED INTO TEXT: On the device, or on a provider's servers?
• WHAT ELSE TRAVELS WITH IT: Transcript text, screen context, account identifiers?
• WHAT SURVIVES PROCESSING: Is anything retained, logged, or used for training?
• DOES IT WORK OFFLINE: Dictate in airplane mode and see what happens.