feat: hot-reload voice models without server restart

Voice enabled/STT model are now DB-backed (admin settings), not env
vars. Added reload_stt_model()/reload_tts_model() that clear singletons
under lock and re-trigger loading. POST /api/admin/voice/reload triggers
both in background tasks. Settings UI polls /api/voice/status every 2.5s
until models are ready, with spinner feedback.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -15,7 +15,7 @@ services:
     environment:
       DATABASE_URL: "postgresql+asyncpg://${POSTGRES_USER:-fabled}:${POSTGRES_PASSWORD:-fabled}@db:5432/${POSTGRES_DB:-fabledassistant}"
       OLLAMA_URL: "http://ollama:11434"
-      OLLAMA_MODEL: "${OLLAMA_MODEL:-llama3.1}"
+      OLLAMA_MODEL: "${OLLAMA_MODEL:-qwen3:8B}"
       SECRET_KEY: "${SECRET_KEY:-dev-secret-change-me}"
       # Uncomment and set to enable web research and image search via SearXNG:
       # SEARXNG_URL: "http://searxng:8080"

frontend/src/views/SettingsView.vue

@@ -450,6 +450,7 @@ const adminVoiceEnabled = ref(false);
 const adminVoiceSttModel = ref("base.en");
 const savingAdminVoice = ref(false);
 const adminVoiceSaved = ref(false);
+const voiceLoadingModels = ref(false);
 
 async function saveAdminVoice() {
   savingAdminVoice.value = true;
@@ -461,13 +462,39 @@ async function saveAdminVoice() {
     });
     adminVoiceSaved.value = true;
     setTimeout(() => { adminVoiceSaved.value = false; }, 2000);
-    // Refresh the user-facing voice status after toggling
     voiceTabLoaded.value = false;
+    if (adminVoiceEnabled.value) {
+      await reloadVoiceModels();
+    }
   } finally {
     savingAdminVoice.value = false;
   }
 }
 
+async function reloadVoiceModels() {
+  voiceLoadingModels.value = true;
+  try {
+    await apiPost("/api/admin/voice/reload", {});
+    // Poll /api/voice/status until both STT and TTS are ready (max 3 min)
+    const deadline = Date.now() + 3 * 60 * 1000;
+    while (Date.now() < deadline) {
+      await new Promise((r) => setTimeout(r, 2500));
+      try {
+        const status = await apiGet<{ enabled: boolean; stt: boolean; tts: boolean }>("/api/voice/status");
+        if (status.stt && status.tts) {
+          voiceStatus.value = status;
+          voiceTabLoaded.value = true;
+          break;
+        }
+      } catch { /* keep polling */ }
+    }
+  } catch {
+    toastStore.show("Failed to trigger model reload", "error");
+  } finally {
+    voiceLoadingModels.value = false;
+  }
+}
+
 // Search test (SearXNG)
 const searxngConfigured = ref(false);
 const searxngUrl = ref("");
@@ -2200,10 +2227,22 @@ FABLE_API_KEY=&lt;your-api-key&gt;</pre>
           <p class="field-hint">Larger models are more accurate but use more RAM and take longer to load.</p>
         </div>
         <div class="actions" style="margin-top:0.85rem;">
-          <button class="btn-save" @click="saveAdminVoice" :disabled="savingAdminVoice">
-            {{ adminVoiceSaved ? 'Saved ✓' : savingAdminVoice ? 'Saving…' : 'Save' }}
+          <button class="btn-save" @click="saveAdminVoice" :disabled="savingAdminVoice || voiceLoadingModels">
+            {{ savingAdminVoice ? 'Saving…' : adminVoiceSaved ? 'Saved ✓' : 'Save' }}
           </button>
-          <span v-if="adminVoiceSaved" class="field-hint">Restart the server to apply model changes.</span>
+          <button
+            v-if="!voiceLoadingModels && adminVoiceEnabled"
+            class="btn-secondary"
+            @click="reloadVoiceModels"
+            :disabled="voiceLoadingModels"
+            title="Reload models without restarting the server"
+          >Reload models</button>
+          <span v-if="voiceLoadingModels" class="field-hint">
+            Loading models… this may take a minute.
+            <span class="voice-admin-spinner">
+              <span></span><span></span><span></span>
+            </span>
+          </span>
         </div>
       </section>
 
@@ -3793,4 +3832,26 @@ FABLE_API_KEY=&lt;your-api-key&gt;</pre>
 .form-actions {
   margin-top: 1rem;
 }
+
+.voice-admin-spinner {
+  display: inline-flex;
+  gap: 3px;
+  align-items: center;
+  margin-left: 0.4rem;
+  vertical-align: middle;
+}
+.voice-admin-spinner span {
+  display: inline-block;
+  width: 4px;
+  height: 4px;
+  border-radius: 50%;
+  background: var(--color-text-muted);
+  animation: va-dot-bounce 1.2s ease-in-out infinite;
+}
+.voice-admin-spinner span:nth-child(2) { animation-delay: 0.2s; }
+.voice-admin-spinner span:nth-child(3) { animation-delay: 0.4s; }
+@keyframes va-dot-bounce {
+  0%, 80%, 100% { transform: scale(0.6); opacity: 0.4; }
+  40% { transform: scale(1); opacity: 1; }
+}
 </style>

src/fabledassistant/routes/admin.py

@@ -228,6 +228,17 @@ async def update_voice_config():
     return jsonify({"status": "ok"})
 
 
+@admin_bp.route("/voice/reload", methods=["POST"])
+@admin_required
+async def reload_voice_models():
+    """Reload STT and TTS models in the background without a server restart."""
+    from fabledassistant.services.stt import reload_stt_model
+    from fabledassistant.services.tts import reload_tts_model
+    asyncio.create_task(reload_stt_model())
+    asyncio.create_task(reload_tts_model())
+    return jsonify({"status": "loading"})
+
+
 @admin_bp.route("/invitations", methods=["POST"])
 @admin_required
 async def create_invite():

src/fabledassistant/services/stt.py

@@ -42,6 +42,15 @@ async def load_stt_model() -> None:
             logger.exception(_load_error)
 
 
+async def reload_stt_model() -> None:
+    """Unload the current model and reload with the current config. Safe to call at runtime."""
+    global _model, _load_error
+    async with _model_lock:
+        _model = None
+        _load_error = None
+    await load_stt_model()
+
+
 def stt_available() -> bool:
     return _model is not None
 

src/fabledassistant/services/tts.py

@@ -56,6 +56,15 @@ async def load_tts_model() -> None:
             logger.exception(_load_error)
 
 
+async def reload_tts_model() -> None:
+    """Unload and reload the TTS pipeline. Safe to call at runtime."""
+    global _pipeline, _load_error
+    async with _pipeline_lock:
+        _pipeline = None
+        _load_error = None
+    await load_tts_model()
+
+
 def tts_available() -> bool:
     return _pipeline is not None