diff --git a/whisrs/config.toml b/whisrs/config.toml
index d5464f1..70aa5d0 100644
--- a/whisrs/config.toml
+++ b/whisrs/config.toml
@@ -15,4 +15,9 @@ audio_feedback = true
 
 [local-whisper]
 model_path = "/home/jeena/.local/share/whisrs/models/ggml-base.en.bin"
-phrase_silence_ms = 250
\ No newline at end of file
+# Batch-like: phrases only flush after 5s of silence, so text lands in one
+# piece right after the key is released (0.1s decode) instead of trickling in
+# while the hold key is still down — injection during a held modifier is
+# unreliable. The 20s force-split still emits during very long continuous
+# speech, so nothing is lost there.
+phrase_silence_ms = 5000
\ No newline at end of file