Fuck You Flow 0.1.0
Hold the right Alt, speak, press it again. The text lands where your cursor is,
in any program. Greek and English, mixed in the same sentence.
Everything runs on your own machine. No account, no subscription, no server, and
no internet connection after setup.
What is in the installer
Nothing is downloaded later. The package carries the speech engine built with
the Vulkan backend, which accelerates on AMD, Intel and NVIDIA cards alike and
falls back to the processor when there is no card; both speech models
(large-v3-q5_0 for accuracy, large-v3-turbo-q5_0 for speed); the Silero
voice activity detector; and the WebView2 bootstrapper for a machine that lacks
it.
The first start reads the machine: the graphics adapters and their memory, the
core count and the RAM. It picks the backend, the model that fits the graphics
memory and the thread count that fits the processor. Settings, Speech models
shows what it found and lets you change any of it.
Speed
On an RTX 3070, from releasing the key to the text appearing: 108 words after
82 s of speech in 0.17 s, 312 words after 152 s in 0.88 s. Each finished phrase
is transcribed while you are still talking, so only the last sentence is left
when you stop.
Engine comparison, same card, same 5 s Greek phrase, large-v3-q5_0: Vulkan
593 ms, CUDA 663 ms, processor alone 12.5 s.
Install
Windows shows a blue "Windows protected your PC" dialog because the installer
carries no purchased signature. Click "More info", then "Run anyway". The
install needs no administrator rights and about 1.7 GB of disk.
SHA256SUMS.txt holds the checksum of the installer.
Known limits
Cards from AMD and Intel are served by the Vulkan build, which was measured on
an NVIDIA card only; reports from other hardware are welcome. A chat box that
locks itself right after a message is sent can drop the paste, in which case the
text stays on the clipboard and Ctrl+V puts it in. The question mark added from
the rise of your voice is tuned on one speaker and will be wrong now and then;
the toggle is in Settings, Transcript cleanup.
See docs/KNOWN-LIMITATIONS.md for the full list.