diff --git a/public/js/clinicalAssistant.js b/public/js/clinicalAssistant.js
index ac1665ac..3c3721d6 100644
--- a/public/js/clinicalAssistant.js
+++ b/public/js/clinicalAssistant.js
@@ -1138,7 +1138,22 @@ import {
     }).catch(function() {});
   }
 
-  var conversationMode = { active: false, listening: false, recorder: null, recognition: null, sessionFinals: '' };
+  var conversationMode = { active: false, listening: false, recorder: null, recognition: null,
+                           sessionFinals: '', speaking: null };
+
+  // Whatever is currently reading an answer aloud, stopped. Without a handle on
+  // the <audio> element there was nothing to stop: ending the call left the
+  // answer talking until the page was reloaded.
+  function conversationStopSpeaking() {
+    try { if (window.speechSynthesis) window.speechSynthesis.cancel(); } catch (e) {}
+    var playing = conversationMode.speaking;
+    conversationMode.speaking = null;
+    if (!playing) return;
+    try { playing.pause(); playing.src = ''; } catch (e) {}
+    if (playing.dataset && playing.dataset.objectUrl) {
+      try { URL.revokeObjectURL(playing.dataset.objectUrl); } catch (e) {}
+    }
+  }
 
   function openConversationMode() {
     if (conversationMode.active) return;
@@ -1179,6 +1194,8 @@ import {
 
   function conversationStartListening() {
     unlockAudioPlayback();
+    // Speaking over the last answer is how a voice assistant talks past you.
+    conversationStopSpeaking();
     var status = document.getElementById('assistant-voice-status');
     if (status) status.textContent = 'Listening…';
     conversationMode.listening = true;
@@ -1246,7 +1263,35 @@ import {
     speakAnswerVoice(String(detail.answer || ''));
   }
 
-  function speakAnswerVoice(text) {
+  // What a person would read out, not what the model wrote. Passed raw, a
+  // browser voice reads "#" and "**" aloud or stumbles over them — the answer
+  // is markdown and the speaker is not a renderer.
+  function speakableText(markdown) {
+    return String(markdown || '')
+      .replace(/```[\s\S]*?```/g, ' ')             // code blocks say nothing useful
+      .replace(/`([^`]+)`/g, '$1')
+      .replace(/!\[[^\]]*\]\([^)]*\)/g, ' ')      // images
+      .replace(/\[([^\]]+)\]\([^)]*\)/g, '$1')    // links keep their words
+      .replace(/^\s{0,3}#{1,6}\s+/gm, '')           // headings are just sentences
+      .replace(/^\s{0,3}>\s?/gm, '')
+      .replace(/^\s*[-*+]\s+/gm, '')                // bullets
+      .replace(/^\s*\d+\.\s+/gm, '')
+      .replace(/^\s*\|.*\|\s*$/gm, ' ')            // a table read aloud is noise
+      .replace(/^\s*[-:|\s]{3,}\s*$/gm, ' ')
+      .replace(/(\*\*|__)(.*?)\1/g, '$2')
+      .replace(/(\*|_)(.*?)\1/g, '$2')
+      .replace(/~~(.*?)~~/g, '$1')
+      .replace(/^\s*[-*_]{3,}\s*$/gm, ' ')          // horizontal rules
+      .replace(/\[\^?\d+\]/g, ' ')                 // citation markers
+      .replace(/[ \t]+/g, ' ')
+      .replace(/ ([.,;:!?])/g, '$1')                // the gap a removed marker leaves
+      .replace(/\n[ \t]*\n+/g, '\n')
+      .replace(/\n{2,}/g, '\n')
+      .trim();
+  }
+
+  function speakAnswerVoice(markdown) {
+    var text = speakableText(markdown);
     var done = function(msg) {
       var st = document.getElementById('assistant-voice-status');
       if (st) st.textContent = msg || 'Tap to speak';
@@ -1256,9 +1301,10 @@ import {
     if (typeof window !== 'undefined' && window.speechSynthesis && typeof SpeechSynthesisUtterance === 'function') {
       try {
         var utter = new SpeechSynthesisUtterance(text);
-        utter.onend = function() { done('Tap to speak'); };
-        utter.onerror = function() { done('Tap to speak'); };
+        utter.onend = function() { conversationMode.speaking = null; done('Tap to speak'); };
+        utter.onerror = function() { conversationMode.speaking = null; done('Tap to speak'); };
         window.speechSynthesis.cancel();
+        conversationMode.speaking = utter;
         window.speechSynthesis.speak(utter);
         return;
       } catch (e) { /* fall through to the model voice */ }
@@ -1271,14 +1317,20 @@ import {
       .then(function(blob) {
         var url = URL.createObjectURL(blob);
         var audio = new Audio(url);
-        audio.onended = function() { URL.revokeObjectURL(url); done(); };
-        audio.onerror = function() { URL.revokeObjectURL(url); done('Voice reply unavailable — tap to speak'); };
-        audio.play().catch(function() { URL.revokeObjectURL(url); done('Voice reply unavailable — tap to speak'); });
+        audio.dataset.objectUrl = url;
+        // The reply arrives after the call may already have ended. Nothing
+        // should start talking into a closed overlay.
+        if (!conversationMode.active) { URL.revokeObjectURL(url); return; }
+        conversationMode.speaking = audio;
+        audio.onended = function() { conversationMode.speaking = null; URL.revokeObjectURL(url); done(); };
+        audio.onerror = function() { conversationMode.speaking = null; URL.revokeObjectURL(url); done('Voice reply unavailable — tap to speak'); };
+        audio.play().catch(function() { conversationMode.speaking = null; URL.revokeObjectURL(url); done('Voice reply unavailable — tap to speak'); });
       }).catch(function() { done('Voice reply unavailable — tap to speak'); });
   }
 
   function endConversationMode() {
     conversationMode.active = false;
+    conversationStopSpeaking();
     conversationStopListening();
     if (activeAssistantRequest) activeAssistantRequest.abort();
     document.removeEventListener('assistant-answer-done', conversationAnswerDone);
diff --git a/test/assistant-voice-mode.test.js b/test/assistant-voice-mode.test.js
new file mode 100644
index 00000000..9ec62fd2
--- /dev/null
+++ b/test/assistant-voice-mode.test.js
@@ -0,0 +1,61 @@
+// ============================================================
+// ASSISTANT VOICE MODE
+// ============================================================
+// The hands-free overlay: tap, speak, hear the answer. Two things about it were
+// wrong in ways a person notices immediately.
+
+const test = require('node:test');
+const assert = require('node:assert');
+const fs = require('fs');
+const path = require('path');
+const read = p => fs.readFileSync(path.join(__dirname, '..', p), 'utf8');
+const src = read('public/js/clinicalAssistant.js');
+
+// The function under test is a closure inside a browser IIFE, so it is pulled
+// out by source rather than imported.
+const speakableText = (() => {
+  const from = src.indexOf('function speakableText');
+  const to = src.indexOf('function speakAnswerVoice');
+  // eslint-disable-next-line no-eval
+  return eval(src.slice(from, to) + '; speakableText');
+})();
+
+test('ending the call stops the voice', () => {
+  // It stopped listening and aborted the request but never stopped playback,
+  // and the <audio> element was a local with no handle kept — so the answer
+  // talked on until the page was reloaded.
+  assert.match(src, /function conversationStopSpeaking\(\)/);
+  assert.match(src, /if \(window\.speechSynthesis\) window\.speechSynthesis\.cancel\(\)/);
+  assert.match(src, /playing\.pause\(\); playing\.src = '';/);
+
+  const end = src.slice(src.indexOf('function endConversationMode'));
+  assert.ok(end.indexOf('conversationStopSpeaking()') < end.indexOf('conversationStopListening()'),
+    'hanging up silences the answer');
+  // Speaking over the previous answer is how an assistant talks past you.
+  assert.match(src, /unlockAudioPlayback\(\);\s*\n\s*\/\/[^\n]*\n\s*conversationStopSpeaking\(\);/);
+  // A reply that arrives after the overlay closed must not start talking.
+  assert.match(src, /if \(!conversationMode\.active\) \{ URL\.revokeObjectURL\(url\); return; \}/);
+  // Both players are tracked, or only one of them can be stopped.
+  assert.match(src, /conversationMode\.speaking = utter;/);
+  assert.match(src, /conversationMode\.speaking = audio;/);
+});
+
+test('the answer is read, not its markup', () => {
+  // The raw answer went straight to the speaker. It is markdown, and a browser
+  // voice reads "#" and "**" aloud or stumbles over them.
+  assert.match(src, /var text = speakableText\(markdown\);/);
+
+  assert.equal(speakableText('## Croup\n\nA **viral** illness.'), 'Croup\nA viral illness.');
+  assert.equal(speakableText('- One\n- Two'), 'One\nTwo');
+  assert.equal(speakableText('1. First\n2. Second'), 'First\nSecond');
+  assert.equal(speakableText('See [the guideline](https://x.y).'), 'See the guideline.');
+  assert.equal(speakableText('Use `epinephrine` now.'), 'Use epinephrine now.');
+  // A table read aloud is noise; the prose around it survives.
+  assert.equal(speakableText('| A | B |\n|---|---|\n| 1 | 2 |\n\nAfter.'), 'After.');
+  // Citation markers leave no gap before the punctuation they preceded.
+  assert.equal(speakableText('Clear [1] and also [^2].'), 'Clear and also.');
+  assert.equal(speakableText('---\n\n> Quoted.'), 'Quoted.');
+  assert.equal(speakableText('```\ncode\n```\n\nThen.'), 'Then.');
+  assert.equal(speakableText(''), '');
+  assert.equal(speakableText(null), '');
+});