Blog

Why local speech-to-text matters

What “local” changes about voice typing, what it does not change, and how SleekVoice’s claim differs from Wispr Flow’s cloud transcription.

Published · Updated · 6 min read · Bharat Bhavnasi

Local speech-to-text means the audio is turned into words on the computer in front of you. The recording does not have to be uploaded to a vendor’s server for a transcript to come back. That is the property people are usually asking about when they say they want private dictation. It is a smaller property than “nobody will ever see what I said.” This post draws that line with SleekVoice’s published behavior on one side and Wispr Flow’s published behavior on the other.

SleekVoice is a free, open-source Mac app. It says Whisper runs locally, including offline, and that your voice stays on your Mac. Wispr’s data-controls page (opens in a new tab) says transcription always occurs on the cloud. Both sentences can be true of their own products at the same time. They describe different places where the vulnerable moment happens: the moment the recording exists as audio, before it is just text in a document.

What leaves the machine in a cloud dictation product

If transcription happens on a server, the audio, or a representation of it, has to get there. Wispr says that path is always cloud, and that this is how they get the latency and accuracy they want. They also describe controls around what happens next.

Dictation Cloud Storage, on Wispr’s privacy page (opens in a new tab) and data-controls page, decides whether transcripts, audio, and history are kept on Wispr’s servers for features that need them on more than one device. Turn it off, Wispr says, and audio is discarded once it is transcribed. That is a retention control. The audio was still processed in the cloud.

A second switch covers whether Wispr may use audio, transcript, and edits to evaluate or improve Wispr’s models. Wispr says third-party model providers are under agreements not to train on that data either way. Wispr’s help center says turning off model improvement does not stop cloud processing, and that some details of a dictation, such as word count, duration, app, and timestamp, can still be uploaded. Read the article before you summarize it as “private mode means offline.” The older name “Privacy Mode,” Wispr says, mapped to the training choice, not to on-device transcription.

Wispr also says context awareness can read limited text from the app you are dictating into, and that you can turn that off. That is a separate exposure: not your voice, but the words already on the screen. SleekVoice has not published an equivalent screen-reading feature. Do not import Wispr’s setting into a description of SleekVoice.

None of this requires believing Wispr is careless. A cloud recognizer can be encrypted in transit, access-controlled, and certified. Wispr’s pricing page says the company has SOC 2 Type II and ISO 27001. This site has not reviewed the reports. Certification is about how a company runs a service. It does not move the recognition back onto the laptop.

What local changes

If the recognizer runs on the Mac, the vendor does not need the audio to produce the transcript, because the vendor’s server is not in the path. SleekVoice’s claim is that version. Whisper on the machine, including with the network off. There is no SleekVoice account whose history would be the copy of record.

You can test the claim in a way that does not require trusting a blog. Install the app, confirm a short phrase appears in a local note, turn the network off, and try again. If the second phrase still appears, that take did not need SleekVoice’s servers. The project does not describe a fallback that uploads audio when the network is available. If a first-run download is required, finish it online, then do the offline check. The project has not named the file that download would fetch.

Local also means a different failure mode. The Mac has to be able to run the model. SleekVoice has not published a chip requirement or a model size, so this post will not invent a thermal or battery number. The tradeoff people usually accept is that the machine does the work, and a vendor’s GPU does not.

The license matters for the same reason the architecture matters. GPL-3.0 source at github.com/SuperAgents-Dev/SleekVoice (opens in a new tab) is how someone other than the author can check whether the app does what the website says, once the source is actually in the tree. A privacy sentence you cannot read in code is still a sentence. An empty repository on a given day does not make the sentence false. It means the check is not available yet. Say that, and look again later.

What local does not change

The transcript is typed into the focused app. From that moment it is ordinary text. If you dictated into Gmail, Slack, Messages, or a document that syncs, those products’ own sync is what moves the words. Local speech-to-text does not wrap the destination in a new encryption layer. It does not stop you from pasting the paragraph into a chatbot. It does not stop a screen share.

It also does not hide the fact that you used a microphone. macOS still has a Microphone permission and, for an app that types into other apps, typically an Accessibility permission. Those switches are visible in System Settings. Apple documents them. They are a record that an app can hear and can type. They are not a copy of your audio on a vendor server.

A personal dictionary of names sits on the Mac as well. Those entries are sensitive in a boring way: they are a list of words you care about spelling. They are not a recording. They are still worth knowing about if you share a Mac user account with someone else. SleekVoice does not describe a per-profile cloud dictionary, because it does not describe a cloud account.

How to ask any dictation vendor the useful questions

You can ask these without a scorecard.

Where is the audio during transcription? “On device,” “on our servers,” or “it depends, and here is the setting” are answers. “We take privacy seriously” is not.

If a server is involved, what is stored after the transcript returns, and what is the default? Retention and training are different switches. Wispr documents them separately. A vendor who only answers one of them has not answered the other.

Does offline mode work, and did you try it? A menu item named offline is weaker than a successful dictation with the network off.

Can a second person read the client? Source is not a privacy proof by itself. It is the artifact that makes the proof possible.

What happens to the text after it is inserted? If the honest answer is “whatever that app does with text,” the vendor understands their own boundary.

SleekVoice’s answers, as published: on the Mac; no SleekVoice cloud history described; offline is part of the claim; source is GPL-3.0 at the repository above; the text lands in the focused app and is then that app’s problem. The privacy page is the short version. The comparison is the Wispr-specific version with links.

Questions

Is local transcription the same as end-to-end encryption?

No. End-to-end encryption is a claim about who can read data in transit or at rest between specific endpoints. Local transcription is a claim about where recognition runs. SleekVoice does not advertise end-to-end encryption.

If I opt out of Wispr model training, is Wispr local?

Wispr says no. Their pages say transcription still occurs on the cloud, and that the training control is separate from Dictation Cloud Storage.

Does SleekVoice send the transcript anywhere?

The project does not describe a transcript upload. The destination app might, after the text is inserted. That is the distinction to keep.

  • privacy
  • speech to text
  • Whisper

Related reading