diff --git a/plugin-designs/rudder-voice/rudder-voice.vue b/plugin-designs/rudder-voice/rudder-voice.vue
index 721c4c750..1cf8e3f61 100644
--- a/plugin-designs/rudder-voice/rudder-voice.vue
+++ b/plugin-designs/rudder-voice/rudder-voice.vue
@@ -124,13 +124,65 @@ async function loadRobotVoices() {
     for (const v of synth.getVoices().filter((x) => x.localService).slice(0, 10)) opts.push({ value: 'system:' + v.voiceURI, label: 'System voice · ' + v.name })
   }
   robotOptions.value = opts
-  if (!opts.some((o) => o.value === robotVoice.value)) robotVoice.value = opts[0] ? opts[0].value : ''
+  if (!opts.some((o) => o.value === robotVoice.value)) {
+    // Default: Male voice 5 (speaker 375) — chosen by ear as the clearest; 1–4 sounded muffled.
+    const preferred = opts.find((o) => /:375$/.test(o.value))
+    robotVoice.value = preferred ? preferred.value : opts[0] ? opts[0].value : ''
+  }
 }
 
 const sayQueue = []
 let saying = false
+// ─── What the robot says: whole sentences, as soon as they're complete, cleaned up ───────────────
+// The engine only sends a FINAL after a pause, so during a minute of continuous talk the robot fell silent
+// and then read everything at once. The model already punctuates as it streams, so a sentence is ready the
+// moment a word ending in . ! or ? is followed by more words: speak it then, without waiting for a pause.
+// Fragments under 3 words ("Okay.") wait and ride along with the next sentence. robotSpoken counts the words
+// of the current utterance already handed to the robot; the final speaks only what's left.
+let robotSpoken = 0
+const SENTENCE_END = /[.!?]["')\]]?$/
+// ★ RUN-ONS (measured 2026-09-28): with no pause between sentences the model joins them with a COMMA, not
+// a period ("…immediately afterwards, going along…"), so during non-stop ideation there may be no sentence
+// end for 20+ seconds. So a long pending stretch also breaks at a clause: ≥14 words waiting → speak up to the
+// last comma (keeping ≥8 words); ≥35 words with no comma at all → speak the first 25, so the robot never
+// falls silent for long.
+const CLAUSE_END = /[,;:]$/
+function robotOnPartial(full) {
+  if (!robotOn.value) return
+  const w = wordsOf(full)
+  const done = w.length - 1            // the last word may be half-streamed: only words before it are settled
+  const pending = done - robotSpoken
+  let end = -1
+  for (let i = robotSpoken; i < done; i++) if (SENTENCE_END.test(w[i])) end = i
+  if (end < 0 && pending >= 14) {
+    for (let i = robotSpoken + 7; i < done; i++) if (CLAUSE_END.test(w[i])) end = i
+    if (end < 0 && pending >= 35) end = robotSpoken + 24
+  }
+  if (end >= robotSpoken && end - robotSpoken + 1 >= 3) {
+    speak(w.slice(robotSpoken, end + 1).join(' '))
+    robotSpoken = end + 1
+  }
+}
+function robotOnFinal(text) {
+  const rest = wordsOf(text).slice(robotSpoken)
+  robotSpoken = 0
+  if (robotOn.value && rest.length) speak(rest.join(' '))
+}
+
+// A light clean-up of what the ROBOT says (the box keeps exactly what was said): drop filler sounds, collapse
+// stuttered repeats ("the the"), tidy spacing, capitalise. Pure text rules, no model, no extra compute.
+// A filler takes its commas with it (", um," → " "), but never a full stop that ends a sentence.
+const FILLERS = /(?:,\s*)?\b(?:u+m+|u+h+m*|e+r+m+|h+m+|a+h+)\b(?:\s*,)?/gi
+function cleanForSpeech(t) {
+  let s = (t || '').replace(FILLERS, ' ')
+  s = s.replace(/\b(\w+)(?:[,\s]+\1\b)+/gi, '$1')
+  s = s.replace(/\s+([,.!?;:])/g, '$1').replace(/([,.!?;:])(?=[^\s\d])/g, '$1 ').replace(/\s{2,}/g, ' ').trim()
+  s = s.replace(/^[,.;:!?\s]+/, '')
+  return s ? s[0].toUpperCase() + s.slice(1) : ''
+}
+
 function speak(text) {
-  const t = (text || '').trim()
+  const t = cleanForSpeech(text)
   if (!t || !robotVoice.value) return
   sayQueue.push(t)
   if (!saying) drainSpeech()
@@ -440,11 +492,12 @@ function onEvent(data) {
       case 'partial':
         partial.value += e.text
         onPartial(partial.value)
+        robotOnPartial(partial.value)
         break
       case 'final':
         partial.value = ''
         onFinal(e.text)
-        if (robotOn.value) speak(e.text)
+        robotOnFinal(e.text)
         break
       case 'recovering':
         say(e.message, 'warn')
@@ -553,6 +606,7 @@ function closeStream() {
 
 // Recording stopped mid-sentence: the streamed words are already in the box, so just let go of them.
 function flushPartial() {
+  robotSpoken = 0
   clearTimeout(flushTimer)
   const finals = pendingFinals.splice(0)
   const p = pendingPartial