--- a/dictation/README.md	Tue Aug 18 22:18:15 2026 -0700
+++ b/dictation/README.md	Thu Aug 20 20:45:25 2026 -0700
@@ -1,8 +1,8 @@
 # WebRTC dictation
 
 This Bazel package runs a local WebRTC speech-to-text service using aiortc and
-faster-whisper. The default multilingual Whisper `small` model uses the RTX
-4070 Ti through native WSL CUDA with `int8_float16` compute.
+faster-whisper. The default multilingual Whisper `small` model uses CUDA with
+`int8_float16` on Linux and CPU inference with `int8` on macOS.
 
 ## Setup
 
@@ -30,6 +30,7 @@
 
 The model is stored under `~/.cache/zenbu/faster-whisper-small` by default.
 Override it with `DICTATION_MODEL_DIR`.
+Set `DICTATION_DEVICE=cpu|cuda` to override platform selection.
 
 Useful tuning:
 
@@ -46,8 +47,8 @@
 and usually a TURN service.
 
 Infinite Canvas hosts this page as a hidden CEF/iframe transport. Pressing `M`
-shows partial and final transcript text in a centered Raylib caption; pressing
-`Enter` sends the accumulated text to Qwen session orchestration:
+shows partial and final transcript text in the retained Dictation entity;
+pressing `Enter` sends the accumulated text to Copilot orchestration:
 
 ```bash
 bazel run //infinite_canvas:agent_dev