diff --git a/server.py b/server.py
index 0ca42d6..bd35eb8 100644
--- a/server.py
+++ b/server.py
@@ -2032,7 +2032,14 @@ def _transcribe_audio(src: Path, settings: dict, backend: str = "configured") ->
                 headers=hdrs,
                 timeout=60,
             )
-    resp.raise_for_status()
+    if not resp.ok:
+        # Extract the real error detail from the response body
+        try:
+            body = resp.json()
+            detail = body.get("detail") or body.get("error") or body.get("message") or str(body)
+        except Exception:
+            detail = resp.text[:300].strip()
+        raise RuntimeError(f"STT backend returned HTTP {resp.status_code}: {detail}")
     return _transcription_text_from_response(resp), backend
 
 
diff --git a/static/app.js b/static/app.js
index ef8f680..a82badc 100644
--- a/static/app.js
+++ b/static/app.js
@@ -8544,7 +8544,14 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
           } else if (evt.type === 'error') {
             const wrap = assistantBubble.closest('.conv-bubble-wrap');
             if (wrap) wrap.remove();
-            addErrorBubble(`[${evt.stage?.toUpperCase() || 'ERR'}] ${evt.message}`);
+            const stage = evt.stage?.toUpperCase() || 'ERR';
+            let msg = evt.message || 'Unknown error';
+            // Surface the actual server error detail, not the raw HTTP noise
+            const detailMatch = msg.match(/HTTP \d+:\s*(.+)/s);
+            if (detailMatch) msg = detailMatch[1].trim();
+            // Truncate very long stack traces
+            if (msg.length > 300) msg = msg.slice(0, 300) + '…';
+            addErrorBubble(`[${stage}] ${msg}`);
             addHistoryItem(turnN, Date.now() - t0, false);
             if (micStatus) micStatus.textContent = 'Error — ready';
           }