Private, on-device text to speech

Most "private" claims don't explain themselves

A lot of text-to-speech apps advertise privacy in a single vague sentence — "your data is safe with us," "100% private" — without saying what actually happens to the text you select. That vagueness is exactly the problem: privacy in this category isn't one fact, it's a pipeline. Where does the text go before it becomes sound? Does it touch a server at all? If it does, who receives it, and is anything kept? Answering those questions precisely is more useful, and more honest, than a slogan.

What "on-device" means for Shmia's free tier

Shmia's free tier uses 10 built-in local voices, and the speech for those voices is generated entirely on your own computer. When you select text and trigger playback — either with the play bubble or the global shortcut — the text is processed locally, converted to audio locally, and played locally. There is no internet round-trip required for the speech itself, because there's no server in that path at all. Practically, that means the actual content of what you're listening to — a personal journal entry, an early draft you're not ready to share, a private email — never gets transmitted anywhere. It stays on the machine you're already trusting with that document in the first place.

What changes if you use premium voices or AI summaries

Shmia's optional Premium ($19.99/month) and Ultra ($39.99/month) tiers add 8 human-sounding ElevenLabs AI voices and AI text summaries. Using either of those features is a genuinely different pipeline: the selected text is sent to Shmia's server, which forwards it to ElevenLabs (for voice generation) or Anthropic (for summaries), and the result streams back to your floating player. Shmia states that it does not store that text — it only counts the usage seconds or summary count against your plan's monthly limit — but it is, unambiguously, a cloud request. That's a meaningfully different privacy posture than the free tier's fully local path, and it should be described as such rather than folded into a blanket "private" claim.

Why this transparency is itself the point

Being specific about which features are local and which are cloud-based isn't a limitation to downplay — it's the actual selling point. It lets you make an informed choice per document: use the free, on-device voices for anything you'd rather keep off a server, and reach for the premium voices only when the specific document doesn't carry that concern and you want a more natural-sounding voice for it. That's a more useful mental model than a single blanket promise, and it's one most competing apps that market themselves as "private" don't spell out this clearly.

The other genuinely offline option: Balabolka

Shmia isn't the only tool that can honestly claim an offline path. Balabolka is free and fully offline — there's no server involved in its speech generation at all, similar in spirit to Shmia's free tier. The tradeoffs are real, though: Balabolka is Windows-only (no Mac version), its built-in voices are the more robotic OS-level kind rather than natural-sounding ones, its interface is dated, and it requires pasting or importing text into the app window rather than a one-shortcut, works-in-any-app mechanic. If you're on Windows only and voice quality and workflow speed matter less to you than having zero moving parts touch a server, it's a fair alternative to consider.