Offline voices with Piper
The local voice is the one that works on a train, on a locked-down laptop and with nothing leaving your PC at all.
Which voice reads is one question under Settings › Voice: Who reads aloud? On Windows the answers are Standard (the voices already in Windows), Natural (Piper voices you download once), Studio (Google, with your own key), OpenRouter (voices from other services, with your own key) and Microsoft Edge. Standard and Natural run entirely on your PC. Natural is the Piper one, and it is what you get after installing.
No voice service ever receives the image, because the text recognition is the one built into Windows or macOS. What differs per service is the sentence text: Studio, OpenRouter and Edge send the sentence to Google, OpenRouter or Microsoft to be spoken. Natural and Standard do not send anything.
Setting up a local voice
Step 1
Choose Natural under Settings › Voice
Open Speakybara and go to Settings › Voice. Under Who reads aloud? pick Natural; after installing it is already chosen. The Windows download comes with two voices, a Dutch and an English one.
On a Mac, pick Local instead: each language then gets a natural voice, one to download, or Automatic with a voice from macOS. The Mac app comes without voices and fetches one when you pick it.
Step 2
Download more voices if you want
Under Voice per language, the row Download voices has a Download button. It opens a list of about 65 more, grouped by language: tick the ones you want and press Download. Each voice file is around 60 MB and downloads once.
Step 3
Pick a voice per language
Every language you have switched on gets its own dropdown under Voice per language, because Speakybara detects the language of your text by itself, per selection and per sentence. The play button next to each one speaks a sample.
Which languages are on is set under Settings › Language and recognition.
Step 4
Set when the voice leaves memory
A Natural voice is loaded into memory while it speaks and unloaded after 90 seconds of silence. Change that under More options at the bottom of the Voice page, Voice out of memory after: longer means the next sentence starts faster, shorter means less memory in use. Unload from memory now does it straight away.
Worth knowing
- None of the voices uses your graphics card: 0 MB of video memory, deliberately.
- With an online voice nothing is loaded locally and the app stays around 55 MB. With a Natural voice, add the voice file while it is in memory.
- If an online service drops out mid-text, Speakybara carries on with the local voice and tells you once, not once per sentence. When the service comes back, it switches back.
- Piper sounds noticeably more synthetic than the Studio or Edge voices. That is the trade for being offline and free.
- One exception to "nothing leaves your PC": with Clean up text before reading it aloud switched on and an online AI, every capture goes to that AI first, even with a Natural voice. The Voice page says so in orange when that is the case.
If something goes wrong
With a Natural or Standard voice, yes, completely: recognising the text and speaking it both happen on your PC. Studio, OpenRouter and Edge need a connection, because the sentence is sent to Google, OpenRouter or Microsoft to be spoken.
That is the voice file being loaded into memory, about 60 MB. Speakybara already starts loading it when you begin a capture, and it stays there until 90 seconds of silence have passed; raise that delay under Settings › Voice › More options if you read in short bursts.
You do not have to choose. Speakybara detects the language per selection and per sentence, and uses the voice you picked for that language. For English words inside a Dutch sentence, see the guide on your own English words.