The voice layer of the creator stack is quietly becoming a privacy product
If you publish across Instagram, TikTok, YouTube, X, LinkedIn, Facebook, Threads, and Pinterest for a living, you already outsource a lot of your production to machines: captions, thumbnails, clip cuts, scheduling. The one step most of us still hand to a server is the human voice — and that is exactly where a launch like Mi Speaks 2.0 is worth your attention. The maker, Blaine Miller, shipped a version that runs neural text-to-speech entirely on an iPhone, with no account and a working airplane mode. For operators who narrate their own content, that is less a novelty than a workflow question: what happens to your repurposing pipeline when voiceover stops being a cloud dependency?
What problem this actually solves — and why “on-device” is the whole story
The 1.x version, by the maker’s own admission, “sounded like a phone reading to you.” It leaned on the system voices, and a fresh iPhone ships only the compact ones. That is a real ceiling for anyone producing faceless YouTube videos, TikTok voiceovers, or audio versions of a newsletter. The 2.0 update adds 28 Studio Voices, American and British, generated by a neural model on the device itself. Miller states plainly that your text is never uploaded, there is no account, and it works in airplane mode. That combination is the pitch.
I have tested a lot of TTS in the creator stack. ElevenLabs, Play.ht, Murf, and the built-in voices in CapCut and Descript all sound good — some of them sound great — but every one of them is a round trip to a server. For most people most of the time, that is fine. Where it stops being fine is when the text you are reading aloud is not public yet: an unreleased script, a client’s embargoed announcement, a paid newsletter draft, a legal document you are reviewing for a brand deal. In my experience, the moment a creator starts narrating pre-publication work, the “where does my text go” question stops being abstract.
Why TikTok creators should care more than LinkedIn ones
The platform economics differ. A LinkedIn post that gets read aloud is a nice-to-have; a TikTok or YouTube Short with a synthetic voice is often the entire product. Faceless channels in the YouTube Partner Program and TikTok’s Creator Rewards ecosystem live and die on narration quality and turnaround speed. If your voice generation runs locally, you can batch a week of scripts on a flight, in a dead zone, or on a locked phone in your pocket — the last one matters because Miller explains that iOS will not let an app use the graphics processor while it is off screen.
That constraint is the most interesting engineering detail in the whole launch. Because the GPU is unavailable in the background, Mi Speaks added a “Prepare Full Document” step: render the whole book up front, then play to the end with the screen locked. If you skip it, Miller says reading does not stop — any passage that is not ready yet falls back to your system voice. I’d bet that fallback is the single biggest source of “why does this sound wrong” complaints, because it is invisible until it happens.
How it stacks up against what you are probably using now
Here is the honest comparison table in my head, and remember this part is my judgment, not a spec sheet.
Cloud TTS services — ElevenLabs, Play.ht, Murf — win on voice variety, emotion control, and API access. If you are generating 200 voiceovers a month for a network of accounts, you want an API and a queue, not a phone app. Nothing in this launch competes with that.
Editing suites — Descript and CapCut — win on being inside the editing timeline. You are not exporting audio and re-importing it; you are fixing a word and moving on.
Reader apps — Speechify and the accessibility features baked into iOS — win on ubiquity and polish at the reading-list use case. They are the incumbents Mi Speaks is really up against, and they are free or cheap.
Where Mi Speaks carves out space is the intersection: neural quality, no upload, no account, offline. That is a narrow but defensible wedge, and it is aimed at a specific person — someone who reads long documents and cares who can see them. The maker’s own framing is telling: he came back to Product Hunt only because the update was “big enough,” and the free tier used to cut you off at 255 characters, which he admits meant nobody could really try it. That is a candid thing to say about your own previous version, and it is the kind of transparency I want from a solo maker.
The features that matter for a repurposing workflow
Beyond the voices, the 2.0 changelog reads like a list of things a content operator would actually ask for. PDFs open as the actual page with the spoken word highlighted where it sits — that is an accessibility win, and Ahmed ZNOUDA flagged it as such in the comments, calling it “a great detail for accessibility.” Chapters are pulled from the book’s outline, or from its headings when it has none. There is an Apple Watch app that pulls the prepared Studio audio from your phone, and a Resume Reading widget.
For repurposing specifically, the audio export in Premium is the feature to watch. If you can render a clean voice track and drop it into CapCut or Descript, a 2,000-word blog post becomes a YouTube narration, a Short, and a podcast episode without a mic in sight. That is the workflow I would actually test, and it is the one the maker does not spend much time selling.
The pricing math, and where it gets murky
The free tier now has no length limit: any document, any voice already on your device, three saved documents. Premium adds the Studio Voices, an unlimited synced library, audio export, and Turbo speeds. You can buy monthly, yearly, or once for life, and subscriptions start with a 7-day trial. The specific prices are not disclosed on the page, which is a gap — for a tool competing against free iOS accessibility features, price is the first thing I would want to see before recommending it to a team.
Where the math breaks
Two things give me pause. First, the “synced library” implies some account or sync mechanism, which sits awkwardly next to “there’s no account.” I’d bet the sync is iCloud-based rather than a Mi Speaks server, but the page does not say, and that is exactly the kind of detail a privacy-first pitch needs to nail down. Second, the GPU limitation means the best experience requires a deliberate prep step. If your workflow is “paste text, hit play, walk away,” you will hit the system-voice fallback more often than you expect — and once a listener hears the compact voice, the illusion of a consistent narrator is gone.
Who this is not for
If you run a social team with shared logins and a review queue, this is not your tool — there is no team tier, no collaboration, no API mentioned. If you need 50 distinct character voices for a multi-host show, cloud services still win. And if you are on Android, this is an iPhone story; the page says nothing about other platforms. Not disclosed, and worth asking before you build a workflow around it.
What I’d watch / test next
Three concrete things an operator can do this week. One: install it and run the same 1,500-word script through three Studio Voices, then listen for where the fallback kicks in — that tells you whether the prep step fits your workflow or fights it. Two: test the audio export against your editing pipeline, because if the exported track drops cleanly into CapCut or Descript, the repurposing math changes immediately. Three: ask the maker directly, in the Product Hunt thread, where the synced library actually lives and whether Android is on the roadmap. My take: the on-device angle is the real story here, not the voice count — and the creators who win the next year will be the ones who notice when a privacy constraint turns into a production advantage.
