diff --git a/docs/guides/medialink.md b/docs/guides/medialink.md
index fdba3332..f3f6394d 100644
--- a/docs/guides/medialink.md
+++ b/docs/guides/medialink.md
@@ -150,7 +150,9 @@ intercom = medialink.Intercom(link, mic, speaker, voice=voice)
 `audiofeatures`, which ignores a fan or road noise that stays on. Elsewhere,
 CircuitPython among them, it uses a level gate: the room's quiet sets a
 floor, and sound 10 dB over it counts. `voice.detector` says which you got.
-The gate can't tell a voice from a door slamming.
+The gate can't tell a voice from a door slamming. Both learn the room's quiet from
+the first sound they hear, and pass over digital silence (frames in which
+every sample is the same), which a microphone can send while it starts.
 
 When both sides speak at once, one gets the floor and the other's words
 are lost until the turn ends, as with two buttons. While the peer talks,
diff --git a/lib/medialink/voice.py b/lib/medialink/voice.py
index 92b1e61f..bb630a99 100644
--- a/lib/medialink/voice.py
+++ b/lib/medialink/voice.py
@@ -15,6 +15,12 @@
 
 Either way a frame counts for ``hold_ms`` after the last one that held
 speech, so the pauses between words don't end a turn.
+
+Both learn the room's quiet from the first frames they hear, so a frame in
+which every sample is the same is passed over: that is a microphone not yet
+running, not a room. An ES7210 on a P4 board sends 100 to 280 ms of it the
+first time it opens after power-up, and a detector that learned the quiet
+from that took the room's own hiss for 3.4 s of speech.
 """
 
 import struct
@@ -62,6 +68,10 @@ def speaking(self):
     def feed(self, pcm):
         """Take a frame of signed 16-bit PCM. True while it is worth sending."""
         self.level = _level(pcm, self.channels)
+        step = 2 * self.channels
+        if bytes(pcm[step:]) == bytes(pcm[:-step]):
+            # Digital silence, from a microphone still starting: not the room.
+            return self.speaking
         if self._frontend is not None:
             voiced = self._voiced_frontend(pcm)
         else:
diff --git a/tests/test_medialink_voice.py b/tests/test_medialink_voice.py
index 6ae5cf14..fc4dae01 100644
--- a/tests/test_medialink_voice.py
+++ b/tests/test_medialink_voice.py
@@ -53,6 +53,23 @@ def test_a_steady_noise_becomes_the_floor(self):
         heard = [voice.feed(f) for f in frames(loud)]
         self.assertFalse(any(heard[-50:]))
 
+    def a_microphone_starting_up(self, detector):
+        """A microphone's first open after power-up, as the P4 board's ES7210
+        gives it: 160 ms of one repeated sample, then 120 ms of zeros, then
+        the room's hiss."""
+        voice = medialink.VoiceActivity(detector=detector, hold_ms=1000)
+        start = frames(struct.pack("<h", -37) * (16 * 160) + bytes(2 * 16 * 120))
+        for frame in start:
+            voice.feed(frame)
+        return [voice.feed(f) for f in frames(hiss(16000 * 3, level=200, seed=7))]
+
+    def test_digital_silence_is_not_the_rooms_quiet(self):
+        self.assertFalse(any(self.a_microphone_starting_up("level")))
+
+    @unittest.skipUnless(HAVE_FEATURES, "needs audiofeatures with voice activity")
+    def test_digital_silence_is_not_the_rooms_quiet_to_audiofeatures(self):
+        self.assertFalse(any(self.a_microphone_starting_up("audiofeatures")))
+
     @unittest.skipUnless(HAVE_FEATURES, "needs audiofeatures with voice activity")
     def test_audiofeatures_is_the_default_where_there_is_one(self):
         voice = medialink.VoiceActivity(hold_ms=0)
@@ -126,7 +143,7 @@ def listen_after_a_click(self, **kwargs):
         again, as it does after each of the peer's turns, and the first
         frame is a click."""
         fmt = Format(16000, 1)
-        click = struct.pack("<h", 12000) * (FRAME // 2)
+        click = struct.pack("<hh", 12000, -12000) * (FRAME // 4)
         sound = click + hiss(16000, seed=5)
         network, a, b = pair()
         voice = medialink.VoiceActivity(detector="level", hold_ms=40)