diff --git a/web/zebra-spaces.html b/web/zebra-spaces.html
index d4a07e0..310bb33 100644
--- a/web/zebra-spaces.html
+++ b/web/zebra-spaces.html
@@ -688,6 +688,9 @@ try {
   <div class="row" id="row-mic-controls">
     <select id="mic-select" title="audio input device — applies once you have the mic"><option value="">input default</option></select>
   </div>
+  <div class="row" id="row-speaker-controls">
+    <select id="speaker-select" title="audio output device — route peers' audio to a specific speaker or DAC"><option value="">output default</option></select>
+  </div>
   <div class="row" id="row-music-mode">
     <label class="note"><input type="checkbox" id="music-mode"> music mode — raw mic, no echo/noise cancellation (for playing audio through it)</label>
   </div>
@@ -1285,6 +1288,7 @@ const ID_KEY = 'zebra-spaces-id-v1';
 const HANDLE_KEY     = 'zebra-spaces-handle-v1';
 const MUSIC_MODE_KEY = 'zebra-spaces-music-mode-v1';
 const MIC_DEV_KEY    = 'zebra-spaces-mic-device-v1';
+const SPK_DEV_KEY    = 'zebra-spaces-spk-device-v1';
 const CAM_DEV_KEY    = 'zebra-spaces-cam-device-v1';
 const THEME_KEY      = 'zebra-theme-v1';
 /* sessionStorage (per-tab) — tracks which call this tab is in so a
@@ -1521,9 +1525,10 @@ $('handle').value = myHandle;
  * last-picked mic + camera deviceIds. Declared up here so the restore
  * runs before their downstream `let` would put them in the temporal
  * dead zone; downstream code now reads from these existing bindings. */
-let musicMode = false, micDeviceId = '', cameraDeviceId = '';
+let musicMode = false, micDeviceId = '', speakerDeviceId = '', cameraDeviceId = '';
 try { musicMode = localStorage.getItem(MUSIC_MODE_KEY) === '1'; } catch(_){}
 try { micDeviceId = localStorage.getItem(MIC_DEV_KEY) || ''; } catch(_){}
+try { speakerDeviceId = localStorage.getItem(SPK_DEV_KEY) || ''; } catch(_){}
 try { cameraDeviceId = localStorage.getItem(CAM_DEV_KEY) || ''; } catch(_){}
 if ($('music-mode')) $('music-mode').checked = musicMode;
 renderIdentity();
@@ -1592,7 +1597,11 @@ const sfuStreamsByPubHex = new Map();   // pubHex -> MediaStream
 
 function attachSfuTrack(uuid, stream){
   let a = remoteAudio.get(uuid);
-  if (!a){ a = document.createElement('audio'); a.autoplay = true; document.body.appendChild(a); remoteAudio.set(uuid, a); }
+  if (!a){
+    a = document.createElement('audio'); a.autoplay = true;
+    document.body.appendChild(a); remoteAudio.set(uuid, a);
+    applySinkTo(a);
+  }
   a.srcObject = stream;
   /* programmatic play in case autoplay policy needs the nudge after a track
    * swap — caller already had a user gesture (entered the space) */
@@ -2152,7 +2161,7 @@ async function sfuPublish(){
   for (const tr of micStream.getTracks()){ tagTrack(tr); pc.addTrack(tr, micStream); }
   setSenderBitrate(pc.getSenders().find(s=>s.track && s.track.kind==='audio'));
   const offer = await pc.createOffer();
-  offer.sdp = preferStereoOpus(offer.sdp, musicMode ? 256000 : 40000, { music: musicMode });
+  offer.sdp = preferStereoOpus(offer.sdp, musicMode ? 510000 : 40000, { music: musicMode });
   await pc.setLocalDescription(offer);
   await waitForIceGathering(pc);
   const res = await fetch(SFU_BASE + '/publish?room=' + encodeURIComponent(roomID) + '&pub=' + myKeys.pubHex, {
@@ -2242,7 +2251,7 @@ async function sfuPublishScreen(){
   const offer = await pc.createOffer();
   /* screen-share audio is always music-grade — system audio capture is
    * what users actually broadcast, not voice */
-  offer.sdp = preferStereoOpus(offer.sdp, 256000, { music: true });
+  offer.sdp = preferStereoOpus(offer.sdp, 510000, { music: true });
   await pc.setLocalDescription(offer);
   await waitForIceGathering(pc);
   const url = SFU_BASE + '/publish?room=' + encodeURIComponent(roomID)
@@ -2257,12 +2266,12 @@ async function sfuPublishScreen(){
   await pc.setRemoteDescription({ type:'answer', sdp: ans.sdp });
   sfuScreenPC = pc; sfuScreenPeerID = ans.peer_id;
   /* raise the RTP-level caps: 6 Mbps for video (high-detail 1080p screen),
-   * 256 kbps for audio (transparent stereo Opus). The codec-level cap was
-   * already raised via preferStereoOpus(). */
+   * 510 kbps for audio (Opus spec ceiling, transparent stereo). The codec-level
+   * cap was already raised via preferStereoOpus(). */
   for (const s of pc.getSenders()){
     if (!s.track) continue;
     if (s.track.kind === 'video') setSenderMaxBitrate(s, 6000000);
-    if (s.track.kind === 'audio') setSenderMaxBitrate(s, 256000);
+    if (s.track.kind === 'audio') setSenderMaxBitrate(s, 510000);
   }
   /* Screen share CANNOT auto-rebuild like mic + camera — getDisplayMedia
    * requires a fresh user gesture (transient activation). Auto-calling
@@ -2334,7 +2343,7 @@ async function sfuPublishGame(iframe){
   }
   stream.getVideoTracks()[0].addEventListener('ended', () => { sfuUnpublishGame(); });
   const offer = await pc.createOffer();
-  offer.sdp = preferStereoOpus(offer.sdp, 256000, { music: true });
+  offer.sdp = preferStereoOpus(offer.sdp, 510000, { music: true });
   await pc.setLocalDescription(offer);
   await waitForIceGathering(pc);
   const url = SFU_BASE + '/publish?room=' + encodeURIComponent(roomID)
@@ -2351,7 +2360,7 @@ async function sfuPublishGame(iframe){
   for (const s of pc.getSenders()){
     if (!s.track) continue;
     if (s.track.kind === 'video') setSenderMaxBitrate(s, 4000000);
-    if (s.track.kind === 'audio') setSenderMaxBitrate(s, 256000);
+    if (s.track.kind === 'audio') setSenderMaxBitrate(s, 510000);
   }
   /* game-share doesn't auto-rebuild on 'failed' — its capture source is
    * a user-selected iframe via Region Capture; the user would have to
@@ -2798,6 +2807,50 @@ async function refreshMicList(){
     else micDeviceId = sel.value;
   } catch(e){ logLine('err','could not list inputs: '+e.message); }
 }
+/* Speaker output picker — routes peer audio to a specific sink (studio
+ * monitors, external DAC, headphones). The codec gains stop at the
+ * <audio> element; this is the one knob that actually upgrades the
+ * playback chain itself.
+ *
+ * setSinkId() needs a non-default deviceId, and labels are gated behind
+ * mic permission (same gUM gate as enumerateDevices for inputs). Safari
+ * lacks setSinkId on HTMLMediaElement entirely — we hide the row in
+ * that case so it doesn't look broken. */
+const spkSupported = (typeof HTMLMediaElement !== 'undefined') &&
+                     ('setSinkId' in HTMLMediaElement.prototype);
+async function refreshSpeakerList(){
+  if (!spkSupported){
+    const row = $('row-speaker-controls'); if (row) row.classList.add('hidden');
+    return;
+  }
+  try {
+    const devs = await navigator.mediaDevices.enumerateDevices();
+    const outs = devs.filter(d=>d.kind==='audiooutput');
+    const sel = $('speaker-select'); if (!sel) return;
+    sel.innerHTML = '';
+    if (!outs.length){ sel.innerHTML = '<option value="">output default</option>'; return; }
+    outs.forEach((o,i)=>{
+      const opt = document.createElement('option');
+      opt.value = o.deviceId; opt.textContent = 'output ' + (o.label || ('speaker '+(i+1)));
+      sel.appendChild(opt);
+    });
+    if (speakerDeviceId && outs.some(o=>o.deviceId===speakerDeviceId)) sel.value = speakerDeviceId;
+    else speakerDeviceId = sel.value;
+  } catch(e){ logLine('err','could not list outputs: '+e.message); }
+}
+/* Apply current speakerDeviceId to a single <audio> element. Safe to call
+ * before the element is in the DOM. Browsers that lack setSinkId silently
+ * skip — we feature-check at the top so this stays cheap on cold paths. */
+async function applySinkTo(el){
+  if (!spkSupported || !el || typeof el.setSinkId !== 'function') return;
+  try { await el.setSinkId(speakerDeviceId || ''); }
+  catch(e){ logLine('err','setSinkId rejected: '+e.message); }
+}
+/* Re-route every live remote-audio element to the newly picked sink. */
+async function applySinkToAll(){
+  if (!spkSupported) return;
+  for (const [, a] of remoteAudio){ await applySinkTo(a); }
+}
 function tagTrack(t){ if (t) t.contentHint = musicMode ? 'music' : 'speech'; }
 /* Firefox sometimes ignores the EC/NS/AGC constraints at getUserMedia time
  * for non-mic sources (e.g. PulseAudio monitor) and applies its default
@@ -2867,9 +2920,11 @@ async function setSenderBitrate(sender){
   try {
     const p = sender.getParameters();
     if (!p.encodings || !p.encodings.length) p.encodings = [{}];
-    /* 256 kbps stereo Opus is roughly transparent for music; 40 kbps mono is
-     * plenty for speech */
-    p.encodings[0].maxBitrate = musicMode ? 256000 : 40000;
+    /* 510 kbps is the Opus spec ceiling — transparent stereo with headroom
+     * for transient-heavy material (orchestral, drums). 40 kbps mono is plenty
+     * for speech. Opus VBR only spends what it needs; idle music stays around
+     * 200–300 kbps and only peaks claim the ceiling. */
+    p.encodings[0].maxBitrate = musicMode ? 510000 : 40000;
     await sender.setParameters(p);
   } catch(_){}
 }
@@ -2887,19 +2942,33 @@ async function setSenderMaxBitrate(sender, bps){
 /* munge the offer SDP so Opus negotiates stereo + a high maxaveragebitrate.
  * Browsers omit stereo=1 unless they're sure the track is stereo, and the
  * codec-level maxaveragebitrate cap (separate from RTP-level maxBitrate)
- * has to be raised explicitly for music to actually use the headroom. */
+ * has to be raised explicitly for music to actually use the headroom.
+ *
+ * Music mode also pins fullband (maxplaybackrate / sprop-maxcapturerate =
+ * 48000) so Opus can't opportunistically narrow to 16/24 kHz under BWE
+ * pressure — and advertises NACK + maxptime:120 so receivers tolerate
+ * larger frames and reactive packet recovery. cbr=0 is explicit so no
+ * browser quirk silently flips us to constant-rate. */
 function preferStereoOpus(sdp, maxAvgBps, opts){
+  const music = !!(opts && opts.music);
   /* DTX is great for voice (silence is silence) but its comfort-noise
    * transitions audibly pop on continuous music signals — keep it OFF
    * in music mode and ON for voice. */
-  const dtx = !(opts && opts.music) ? '1' : '0';
-  return sdp.replace(/a=fmtp:(\d+) ([^\r\n]*minptime=10[^\r\n]*)/g, (m, pt, fmtp) => {
+  const dtx = music ? '0' : '1';
+  let opusPT = null;
+  sdp = sdp.replace(/a=fmtp:(\d+) ([^\r\n]*minptime=10[^\r\n]*)/g, (m, pt, fmtp) => {
+    opusPT = pt;
     /* useinbandfec=1: forward error correction so a single dropped
-     *   packet doesn't audibly chop — Opus reconstructs from FEC. */
+     *   packet doesn't audibly chop — Opus reconstructs from FEC.
+     * maxplaybackrate / sprop-maxcapturerate=48000: fullband both ways.
+     * cbr=0: explicit VBR. Opus VBR only spends what it needs; the high
+     *   ceiling is just headroom for transients, not a constant load. */
     const want = {
       'stereo': '1', 'sprop-stereo': '1',
       'maxaveragebitrate': String(maxAvgBps),
       'useinbandfec': '1', 'usedtx': dtx,
+      'maxplaybackrate': '48000', 'sprop-maxcapturerate': '48000',
+      'cbr': '0',
     };
     const parts = fmtp.split(';').map(s => s.trim()).filter(Boolean);
     const seen = new Set();
@@ -2911,6 +2980,23 @@ function preferStereoOpus(sdp, maxAvgBps, opts){
     for (const k of Object.keys(want)) if (!seen.has(k)) parts.push(k + '=' + want[k]);
     return 'a=fmtp:' + pt + ' ' + parts.join(';');
   });
+  if (!opusPT) return sdp;
+  /* Add NACK feedback for Opus if not already present — reactive packet
+   * recovery, near-zero overhead. Browsers that don't honor it will
+   * ignore the line. */
+  const nackLine = 'a=rtcp-fb:' + opusPT + ' nack';
+  if (sdp.indexOf(nackLine) === -1){
+    sdp = sdp.replace(new RegExp('(a=rtpmap:' + opusPT + ' opus[^\\r\\n]*\\r?\\n)'),
+                      '$1' + nackLine + '\r\n');
+  }
+  /* Tolerate larger frames from peers (up to 120ms). Bigger encode
+   * windows give Opus more context per packet → cleaner music at the
+   * same bitrate. We can't force OUR encoder's frame size from JS but
+   * advertising the tolerance keeps the negotiation open. */
+  if (music && !/a=maxptime:/.test(sdp)){
+    sdp = sdp.replace(/(a=rtpmap:\d+ opus[^\r\n]*\r?\n)/, '$1a=maxptime:120\r\n');
+  }
+  return sdp;
 }
 async function applyMicMode(){
   /* re-acquire mic with new constraints, hot-swap onto every live sender
@@ -3584,7 +3670,7 @@ async function ensureMicAndUI(){
    * mic and unblock mute. Toggling music-mode before getting a mic is fine —
    * micConstraints() reads the live `musicMode` flag whenever we re-acquire. */
   try {
-    await getMic(); await refreshMicList();
+    await getMic(); await refreshMicList(); await refreshSpeakerList();
     $('btn-mute').disabled = false;
     /* restore the user's last mute choice. Hard refresh keeps
      * sessionStorage so reconnecting in muted state is preserved
@@ -3656,7 +3742,11 @@ async function connectToPeer(uuid, weOffer){
   setSenderBitrate(pc.getSenders().find(s=>s.track && s.track.kind==='audio'));
   pc.ontrack = (ev) => {
     let a = remoteAudio.get(uuid);
-    if (!a){ a = document.createElement('audio'); a.autoplay = true; document.body.appendChild(a); remoteAudio.set(uuid, a); }
+    if (!a){
+      a = document.createElement('audio'); a.autoplay = true;
+      document.body.appendChild(a); remoteAudio.set(uuid, a);
+      applySinkTo(a);
+    }
     a.srcObject = ev.streams[0] || new MediaStream([ev.track]);
     /* 700ms jitter buffer — matches the SFU path. Cross-Wi-Fi and
      * cellular hand-off jitter ate into the old 400ms cushion and
@@ -3702,7 +3792,7 @@ async function connectToPeer(uuid, weOffer){
   };
   if (weOffer){
     const offer = await pc.createOffer();
-    offer.sdp = preferStereoOpus(offer.sdp, musicMode ? 256000 : 40000, { music: musicMode });
+    offer.sdp = preferStereoOpus(offer.sdp, musicMode ? 510000 : 40000, { music: musicMode });
     await pc.setLocalDescription(offer);
     await waitForIceGathering(pc);
     await sendEncSDP(uuid, 'offer', pc.localDescription);
@@ -4143,6 +4233,14 @@ $('mic-select').addEventListener('change', async (e) => {
   try { localStorage.setItem(MIC_DEV_KEY, micDeviceId); } catch(_){}
   if (micStream){ try { await applyMicMode(); } catch(err){ logLine('err','input switch failed: '+err.message); await refreshMicList(); } }
 });
+if ($('speaker-select')){
+  $('speaker-select').addEventListener('change', async (e) => {
+    speakerDeviceId = e.target.value;
+    try { localStorage.setItem(SPK_DEV_KEY, speakerDeviceId); } catch(_){}
+    await applySinkToAll();
+    logLine('', 'output: '+(e.target.selectedOptions[0]?.textContent || 'default'));
+  });
+}
 $('music-mode').addEventListener('change', async (e) => {
   musicMode = e.target.checked;
   try { localStorage.setItem(MUSIC_MODE_KEY, musicMode ? '1' : '0'); } catch(_){}
@@ -4152,6 +4250,7 @@ $('music-mode').addEventListener('change', async (e) => {
 if (navigator.mediaDevices && navigator.mediaDevices.addEventListener){
   navigator.mediaDevices.addEventListener('devicechange', () => {
     refreshMicList();
+    refreshSpeakerList();
     /* belt-and-suspenders for browsers that don't fire track.onended on
      * device disappearance (Firefox/BT swap can leave the track in
      * 'live' state but emitting silence). If our current track has
@@ -4169,6 +4268,7 @@ if (navigator.mediaDevices && navigator.mediaDevices.addEventListener){
  * until mic permission is granted; deviceIds still populate so the user
  * sees how many inputs exist. */
 refreshMicList();
+refreshSpeakerList();
 
 /* tab-close strong-leave: when the page is about to unload, send 'bye'
  * BEFORE the WS gets torn down by the browser. Without this, closing a
@@ -4317,8 +4417,8 @@ logLine('', 'ready — pick a handle, type a rendezvous code, enter the space');
 
 <footer style="margin:2.2rem auto 0;font-size:0.65rem;color:#999;line-height:1.7;word-break:break-all;font-family:monospace">
   <span id="pi-seal" style="color:#777;cursor:default;user-select:none" title="">page integrity</span> &nbsp;·&nbsp; built <span class="stamp-date">2026-06-03</span><br>
-  md5 <span class="stamp-md5">06817aa05772655ad595a070eb160055</span><br>
-  sha256 <span class="stamp-sha">5ddf2dc0fb7e2ecf91e3c0c0b88b9ccd386d068c74df5975c0f509803352164c</span><br>
+  md5 <span class="stamp-md5">45650495593e448837687d642021f48c</span><br>
+  sha256 <span class="stamp-sha">0a00722851d1749740821391946ccf77e8172ae5c024f564b8897fb05e020bc6</span><br>
   <span style="color:#bbb">hashes are of this page with these two fields zeroed — to verify, blank them and re-hash</span><br>
   <span style="color:#bbb">one self-contained file — <strong>save a copy</strong> and verify against these hashes; point at your own servers with ?signal= and ?turncred=, or <a href="host-your-own.html" style="color:#999">host your own community</a></span>
 </footer>