Most announcement systems that speak typed text send that text to a server. It works, it sounds good, and it introduces three problems a building would rather not have: it stops when the internet does, it bills per character, and every announcement you make is somebody else's log entry.
Running the voice on the machine removes all three. Here is what it costs in return, stated honestly.
Running the voice locally matters for four reasons, and only the first is about the internet being unreliable. It keeps working during an outage, it costs nothing per announcement, the text of your announcements stays inside the building, and there is no account or API key that can expire, change price or be deprecated out from under a system somebody installed three years ago.
Speed. The model does real work, on the machine, and this is what it measures at: on a Raspberry Pi 400 the engine synthesises at a real-time factor of roughly 2.1, and time-to-first-audio is a handful of seconds cold and about four seconds for a pre-warmed phrase. On a desktop PC it is far quicker. Figures, hardware and method are published on the measured data page.
The practical consequence: scheduled announcements are pre-warmed and start on time; typed live announcements have a short pause between pressing play and the first word. That pause is the honest price of not using a cloud.
Disk. The voice model is an ordinary file that ships with the software. It is measured in tens of megabytes, not gigabytes, which is the reason it runs on a Pi at all.
Language. This is the real limitation, and it deserves plain words rather than a footnote: the built-in neural voice speaks English only, in US and UK accents. It does not speak Arabic, Urdu, Bengali, French or anything else, and no amount of configuration will make it. Sites that need announcements in another language schedule recorded audio files, which the scheduler treats identically to spoken text in every respect.
Yes, and you will want to. Place names, surnames and abbreviations are where every speech engine embarrasses itself. There is a pronunciation editor built in: type the word, type how it should sound, hear it immediately, and it applies everywhere from then on — schedules, live announcements, the lot. The dictionary is yours, stored with your configuration.
You do not have to wait until you hear a mistake, either. The Pronunciation Workbench lists every unique word the system can say and marks the ones the voice is only guessing at, so you can find the awkward names before a congregation does. And if a word is easier to say than to respell, record yourself saying it once or twice: the device works out the pronunciation from your recording and offers you candidates to audition. Your recording is only ever the target — it is never played in an announcement, so everything stays in one consistent voice.
Hosted speech is genuinely better in three places, and pretending otherwise would be silly. It covers languages other than English, which is not a close call; it has more range on long or expressive passages, because the model is far larger; and it works on hardware too small to run any model at all, where a recorded file is the sensible answer anyway. In detail:
The engine and the model are used under permissive licensing, deliberately chosen so that shipping them imposes no copyleft obligation on anyone deploying the software. This matters more than it sounds: a great deal of open speech tooling is GPL-licensed, which is fine for a hobby install and a genuine problem for an institution that has been told to keep its configuration and integrations private.
Does any part of the announcement reach the internet?
No. Synthesis happens on the device. The only text that leaves is text you deliberately send — for example an announcement issued through the optional cloud relay from outside the building.
How good does it actually sound?
Neural, not robotic. It is a small model, so it is not indistinguishable from a person on a long passage — but for the sentence lengths a building actually uses, most listeners do not notice it is synthetic.
Can I use my own recorded voice instead?
Yes, anywhere. Recorded files and typed speech are scheduled the same way, and many sites use recordings for the handful of announcements that repeat daily and typed speech for everything ad hoc.
Does it work on a Raspberry Pi Zero 2 W?
It runs. It is slower than a Pi 4, which shows up as a longer pause before a live announcement starts; scheduled announcements are unaffected because they are prepared in advance.
Can I change the voice?
Yes — several English voices ship, and the speaking rate is adjustable. Pick one and stay with it: consistency matters more than preference in a building where people learn to recognise the announcement voice.