Local and cloud text-to-speech can use similar voices while handling data very differently. The right choice depends on the sensitivity of the script, the device, and the reliability you need.
This guide separates the stages people often blur: importing a file, editing text, generating speech, storing audio, and analytics. A local import does not automatically mean local speech.
What local processing means
On-device mode downloads a speech model into the browser and processes the script there. After the model is available, text does not need to be submitted to a speech server. The trade-offs are a larger first download, higher memory use, slower performance on older hardware, and varying browser support.
Local mode is a privacy control, not a guarantee that nothing else on the page talks to the network. Ads, analytics, and account requests can still exist on a site. Read the privacy notice for the product you are using. On FreeTextoSpeech, cloud text is processed to return audio and is not kept as a library of your scripts; on-device generation keeps the script on the device after the model is present. Details are in the privacy policy.
Local also fails differently. If the model cannot load, the device is low on memory, or the browser is unsupported, you will notice immediately. That is preferable to a silent leak, but it is still a failure mode you should plan for.
What cloud processing means
Cloud mode sends the selected text to a hosted speech service and returns audio. It usually starts faster on low-powered devices and can provide predictable performance, but the text leaves the device for processing. Read the privacy notice, retention policy, and provider terms before submitting confidential material.
Cloud is often the right tool for public marketing copy, ordinary creator scripts, and machines that cannot run a local model. It is the wrong default for unpublished client work, health notes, legal drafts, and anything your contract says must stay inside a controlled environment.
Treat the character allowance and shared-key limits as operational facts, not as a privacy feature. A quota is not a data-processing agreement.
Choose by risk
Use local mode for unpublished client scripts, personal documents, medical or legal notes, and material covered by an internal confidentiality rule. Cloud mode may be appropriate for public marketing copy, ordinary creator scripts, or devices that cannot run a local model.
A short decision list:
- Would you paste this into a random web form? If no, do not send it to cloud speech.
- Is the document already public? Cloud is usually fine.
- Does a contract, school rule, or regulation restrict processors? Prefer on-device, or an environment your counsel has approved.
- Is the device a shared library computer? Local models can remain on that browser profile. Know who uses the machine.
- Do you need the audio tomorrow on a phone that struggles with large models? Plan a private machine for generation, then carry the file.
Verify the whole workflow
Local document extraction does not automatically mean local voice generation. Check each stage: import, text editing, speech generation, storage, analytics, and download. FreeTextoSpeech labels its on-device and cloud modes so the choice remains explicit.
Worked examples:
- Read PDF Aloud parses the PDF in the browser. Cloud speech still sends the selected text unless you choose the in-browser engine.
- EPUB to speech reads the spine locally, then generates per chapter with the engine you select.
- The homepage cloud tool sends the text you submit so it can return audio. It is not a local reader.
After generation, the downloaded WAV or MP3 is a copy you control. Where you put that file — desktop, private drive, public cloud folder — is a second privacy decision. Speech generation cannot protect a file you later email to the wrong person.
Practical safeguards
- Prefer on-device for anything you would not say aloud in an open office.
- Generate on a machine you control, then move the audio with intent.
- Do not paste secrets, passwords, or one-time codes into any speech form.
- Keep glossaries of names in the project folder, not in a public ticket.
- Re-read the on-screen privacy notice when the product offers a mode switch. Modes change.
Privacy is a workflow property. The same voice can be used safely or carelessly depending on which engine processed the text.
