Can Superwhisper S1-mini Keep Voice Dictation Local? The Audio Catch

August 23, 2026

Superwhisper S1-mini can clean transcripts on-device, but it does not transcribe audio. Check the voice model, history, clipboard, and destination app too.

Superwhisper S1-mini can keep one important part of voice dictation local, but it cannot prove the whole dictation stayed on your device. S1-mini is a small language model that cleans an existing transcript. It does not listen to your microphone or convert audio into words. The voice model that creates the raw transcript can still be local or cloud-hosted, and the resulting audio, text, history, clipboard, and destination app each have separate privacy boundaries.

The practical answer is simple: S1-mini makes local transcript cleanup possible. To claim local voice dictation, you also need a local speech-recognition model and a controlled storage and delivery path.

What Is Confirmed

Superwhisper announced the S1 model family on August 19, 2026. The family includes three different products with different data paths:

• S1-mini is an on-device text-cleanup model.
• S1-Voice is a cloud-hosted speech-to-text model.
• S1-Language is a cloud-hosted model for more advanced cleanup, formatting, and summarization.

That distinction matters more than the shared S1 name. Superwhisper says S1-mini performs its cleanup locally with no network request for that step. Its August 19 changelog likewise describes S1-mini as an on-device language model that formats dictated text without a network request.

The current S1-mini model card on Hugging Face is even more specific. It calls the model a text normalizer for speech-to-text output, not an automatic speech-recognition model. The model receives a raw transcript and rewrites it as cleaner text by removing fillers, resolving false starts, applying punctuation and capitalization, and formatting items such as dates, currency, and email addresses.

At the time of this review, the model card reported:

The model card also shows why the word cleanup is essential. Its documented pipeline is:

audio -> speech recognition -> raw transcript -> S1-mini -> cleaned text

S1-mini begins after speech recognition has already happened. It cannot determine whether the audio was transcribed by a local model, Superwhisper's S1-Voice cloud model, or another provider.

Superwhisper's sensitive-data guide confirms that the product has two independently configurable processing stages. Stage one converts voice to raw text. Stage two optionally processes that text with a language model. Either stage may be local or cloud-based, depending on the selected models and platform.

What Is Still Unclear

The release does not establish that every S1-mini session is end-to-end local by default, nor does it independently validate every performance or privacy claim.

Important unresolved or configuration-dependent questions include:

• Which voice model is active? S1-mini can receive a transcript produced locally or in the cloud.
• Which platform is being used? Superwhisper's current sensitive-data guide says local language models are supported on macOS, while its Windows documentation says local language-model support is not yet available there.
• What does the application save? Superwhisper documents local history containing recordings and transcription data. Local is a meaningful boundary, but it is still stored data.
• What other text enters processing? Dictation tools may use selected text, custom vocabulary, the clipboard, or active-app context in addition to the spoken words.
• Where does the result go? A local transcript pasted into a cloud email client, messaging service, AI chatbot, or collaborative document is no longer confined to the device.
• How broad is independent validation? The public accuracy figures are vendor-reported. This review found current community discussion and user timing reports, but not an independent privacy audit or broad replication of the full benchmark.
• What changes after an update? A model selection, suggested default, operating-system release, or application update can change the effective path even when the mode name stays the same.

None of those unknowns contradicts the documented local execution of S1-mini itself. They show why a model-level fact should not be promoted into an application-wide privacy claim without checking the surrounding system.

The Audio Catch: S1-mini Does Not Hear Your Voice

The easiest mistake is to see a speech-to-text product announce an on-device model and assume that the model handles the microphone audio. S1-mini does not.

Automatic speech recognition and transcript normalization solve different problems:

1. Speech recognition turns an audio waveform into words.
2. Transcript normalization turns rough words into readable writing.

If you dictate, "send it Friday—no, Thursday," the voice model first has to recognize those sounds. S1-mini can then remove the false start and produce a clean sentence ending in Thursday. The second operation can run locally even if the first operation required a cloud request.

Superwhisper's own launch guidance recommends two example combinations:

• An offline combination using a local voice model with S1-mini.
• A cloud-hosted combination using S1-Voice with S1-Language.

Those examples make the architectural choice visible. A user who wants audio to remain on-device must select a local voice model as well as S1-mini. Selecting only S1-mini proves nothing about where the microphone recording was transcribed.

This two-stage distinction is useful beyond one app. Any voice workflow may contain several independently hosted components: recording, noise reduction, speech recognition, language-model cleanup, summarization, storage, synchronization, and delivery. A local model in one box does not relocate the other boxes.

Local Processing Is Not The Same As No Saved Copy

Network privacy and storage privacy are separate questions.

Superwhisper's current history documentation says users can view prior dictations, reveal their files in Finder, and delete recordings. It also says past recordings are not automatically deleted. The separate history-management guide describes local history files containing audio recordings and associated transcription data.

That does not mean the files were uploaded. It means a person assessing privacy should ask two different questions:

• Did the audio or text leave the device during processing?
• What audio or text remains on the device after processing?

A locally stored recording can still be exposed through device theft, another signed-in user, overly broad filesystem permissions, endpoint malware, cloud backup software, a synced Documents folder, or an accidental support attachment. Full-disk encryption, a locked user account, sensible backups, and a deliberate retention period still matter.

The same principle applies to S1-mini model files. Downloading open weights and running inference locally reduces dependence on a remote inference service. It does not encrypt the transcript, secure the operating system, delete history, or control the application that receives the final text.

Use The VOICE Test Before Calling A Dictation Workflow Local

The VOICE test turns a vague privacy label into five observable checks.

V — Voice Model

Identify the model that converts audio into the raw transcript. Do not stop at the language model used for cleanup.

Ask:

• Is the voice model labeled local or cloud?
• Is its model file downloaded to the device?
• Does the same mode behave differently on macOS, Windows, and iOS?
• Does changing languages silently switch to a cloud-only model?

For S1-mini, this is the decisive first check. The cleanup model cannot make a cloud voice model local.

O — Outbound Requests

Observe the network path during a harmless test. A product label is useful evidence, but a controlled network test can reveal unexpected dependencies.

Run a synthetic dictation with no real personal information. Then repeat it while disconnected from the internet or while monitoring outbound connections. Check whether transcription and cleanup still complete, whether the app reports a clear error, and whether any feature quietly falls back to another model.

An offline success is strong evidence for that tested path. It is not proof about every mode, language, update, analytics event, or destination application.

I — In-Device History

Find every local copy created by the workflow:

• Original audio
• Raw transcript
• Cleaned transcript
• Application history
• Temporary files or caches
• Clipboard contents
• Backups and synchronized folders

Set a retention rule based on the sensitivity of the material. If the app does not automatically expire history, periodic deletion may be necessary. Preview exact deletion targets before using any automated cleanup command.

C — Context And Clipboard

Check what else the application can read and where the final text is delivered.

A voice tool may incorporate highlighted text, an active window, a custom vocabulary, or clipboard state. It may paste the result into another application or simulate keystrokes. Superwhisper's advanced-settings documentation describes clipboard and text-delivery controls, which means the destination is part of the privacy model.

If S1-mini produces text locally and that text is immediately pasted into a consumer AI chat, the destination service's retention and training controls apply to the pasted copy. Local preprocessing does not grant local treatment to the next service.

E — Evidence And Updates

Record the exact app version, operating system, mode, voice model, language model, and relevant settings that passed your test. Recheck after changing any of them.

The August 19 release added S1-mini, and the August 20 update included an Intel Mac stability improvement. That is ordinary product evolution, but it illustrates why privacy reviews should attach to a tested configuration and date rather than a permanent brand-level assumption.

A Safe Test You Can Run Without Exposing Real Data

Use a synthetic phrase that is unique enough to trace but meaningless outside the test, such as:

Project Juniper's sample code is 4827. Replace Friday with Thursday and format the result as an email.

Then run this sequence:

1. Select a clearly labeled local voice model.
2. Select S1-mini as the cleanup model.
3. Close unrelated applications that might receive the text.
4. Dictate the synthetic phrase once while connected.
5. Confirm that the raw and cleaned outputs match the expected transformation.
6. Disconnect the network and repeat the test.
7. Inspect the application's local history and recording folder.
8. Check the clipboard and destination application for extra copies.
9. Delete the test artifacts using documented controls.
10. Repeat after an app update, model change, or operating-system change.

Do not use a real password, health record, client name, financial identifier, or confidential project detail as a probe. The goal is to map the path, not to gamble sensitive information on an unverified configuration.

What This Release Does Not Prove

S1-mini's local execution does not prove that:

• Every Superwhisper voice model runs locally.
• Every Superwhisper platform supports local language-model cleanup.
• No diagnostic or account request ever reaches Superwhisper infrastructure.
• Local recordings are automatically deleted.
• A device backup does not copy the recording elsewhere.
• The clipboard or destination app keeps the text private.
• The model's output is always accurate.
• Open weights make the full application open source.
• A particular configuration satisfies HIPAA, GDPR, or another legal requirement.

The vendor's sensitive-data guide itself says compliance depends on the selected configuration, agreements, and organizational controls. That is the right way to state the boundary: a technical option can support a privacy or compliance program without certifying the entire program.

What This Means For OpenVeil

S1-mini is a strong example of the control available from a genuinely local processing stage. If fully local voice processing is your highest priority and you are comfortable managing local models, a correctly configured local dictation stack may be the better fit.

OpenVeil serves a different need. It is a privacy-focused hosted AI workspace with voice tools, browser-local normal chat history, and no normal server-side chat-history record for private chat sessions. It is designed for adults who want chat, web search, files, voice, images, video, and custom personas without maintaining a local AI stack.

OpenVeil is not fully offline. Active voice, transcript, chat, and media requests may still be processed by OpenVeil and necessary providers. OpenVeil also does not inspect Superwhisper, change its model settings, delete its local recordings, or control a destination app.

That tradeoff is the useful comparison:

Before choosing, compare the broader private AI chat versus local AI tradeoff, review how voice, files, and images can follow different privacy settings, and read the OpenVeil privacy policy.

Frequently Asked Questions

Is S1-mini a speech-to-text model?

No. S1-mini is a text-normalization model. It receives a raw transcript created by a separate automatic speech-recognition model and rewrites that transcript into cleaner text.

Does S1-mini send transcripts to the cloud?

Superwhisper says S1-mini runs on-device without a network request for its cleanup stage. That statement does not cover the separate voice model, application account traffic, the destination app, or other modes that use cloud-hosted models.

Can Superwhisper run fully offline?

Superwhisper documents configurations that combine a local voice model with a local language model. Platform and model availability matter. Test the exact mode offline and verify that no cloud fallback is active before relying on the label.

Does local dictation mean no recordings are saved?

No. Local processing describes where computation occurs. Superwhisper's documentation says dictation history, including recordings and associated transcription data, is stored locally and is not automatically deleted. Users can manage those files and choose a retention practice.

Is S1-mini open source?

Superwhisper describes S1-mini as an open-weight release. The model card provides downloadable weights and a license based on Apache 2.0 with an additional naming requirement. Open weights do not by themselves make the surrounding Superwhisper application open source.

Can S1-mini run on an ordinary laptop?

The model card says the 462 MiB quantized build runs comfortably on a laptop CPU. Performance still depends on the device, runtime, input length, and model format.

Does S1-mini support every language?

No. The current v1 model card describes S1-mini as English-only. A different language choice may require another local model, cloud processing, or no S1-mini cleanup.

Is a local voice model automatically safer than a hosted one?

It reduces network exposure for the tested inference path, but it does not solve device security, local retention, backups, malware, clipboard exposure, or destination-app handling. Privacy depends on the full workflow.

Is OpenVeil a fully local alternative to Superwhisper?

No. OpenVeil is a hosted privacy-focused AI workspace. It keeps normal chat history in the browser and does not maintain a normal server-side chat-history record, but active requests still require processing by OpenVeil and necessary providers.

Bottom Line

S1-mini can keep transcript cleanup local. It cannot, by itself, keep voice dictation local because it never handles the audio-to-text step. To make the stronger claim, verify a local voice model, no unexpected outbound request, deliberate local-history retention, controlled clipboard and context access, and a destination app whose data practices match your needs.

The durable lesson is bigger than Superwhisper: privacy follows the whole data path, not the smallest model in the pipeline.