docs: correct whisrs dependencies and describe hold-to-talk

wtype is not actually used by whisrs: injection happens through the
Wayland virtual-keyboard protocol with a uinput fallback, so listing
wtype as a dependency would send fresh installs down the wrong path.
pipewire-alsa on the other hand is essential and was missing: whisrs
captures through raw ALSA, and without the PipeWire bridge the default
PCM is the silent headset-jack mic.

Also document the hold-to-talk binding, the [audio] device setting, and
the local rebuild (-march=native plus the vulkan feature) that moved
decoding to the iGPU, since the stock AUR build is far too slow for
dictation.

README.md

@@ -106,7 +106,10 @@ You also need to install all the dependencies:
   portal startup script falls back to launching the portals manually.)
 - fcitx5
 - fcitx5-hangul
-- wtype (types the transcribed text)
+- pipewire-alsa (whisrs captures audio via raw ALSA; without it the default
+  PCM is the headset-jack mic, which is silent when nothing is plugged in)
+- vulkan-intel and vulkan-icd-loader (iGPU decode; already standard on systems
+  with Intel graphics set up for Vulkan)
 - whisrs-git (AUR; voice dictation daemon, see below)
 
 There are probably more, I will update the list next time I'm
@@ -116,15 +119,25 @@ Voice dictation
 ===============
 
 The Copilot key (on newer laptops this replaced the right Ctrl key; it
-sends SUPER+SHIFT+F23) toggles voice dictation. Press it once to start
-recording and again to stop. whisrs transcribes locally with a
-whisper.cpp model and types the text at the cursor via wtype, in any
-window, including terminals running the dockerized agent harnesses.
+sends SUPER+SHIFT+F23) toggles voice dictation. Hold it to record and
+release to transcribe: one whisrs toggle fires on press, another on
+release. whisrs transcribes locally with a whisper.cpp model and types
+the text at the cursor in any window, including terminals running the
+dockerized agent harnesses. Text injection uses the Wayland
+virtual-keyboard protocol on Hyprland, with a uinput fallback that the
+package's udev rule grants access to.
 
-The config lives in ~/.config/whisrs/config.toml and the model in
+The config lives in ~/.config/whisrs/config.toml (set [audio] device =
+"pipewire" so capture goes through PipeWire) and the model in
 ~/.local/share/whisrs/models/. The daemon runs as the systemd user
 service whisrs.service.
 
+The stock AUR build decodes on the CPU with generic -march flags, which
+makes dictation take around five seconds. On this machine the package
+is rebuilt locally with -march=native plus the vulkan cargo feature so
+decoding runs on the Intel iGPU instead; that drops latency to roughly
+a second. Re-apply those two PKGBUILD edits after AUR updates.
+
 License
 =======