diff --git a/src/servers/api/server-settings/tts.ts b/src/servers/api/server-settings/tts.ts
index 66ec434d..3be5121d 100644
--- a/src/servers/api/server-settings/tts.ts
+++ b/src/servers/api/server-settings/tts.ts
@@ -41,6 +41,24 @@ ttsRouter.put('/', async (ctx) => {
   return ctx.json({ success: true });
 });
 
+// OpenAI-compatible TTS servers disagree on this payload: some list voices as plain strings, others
+// (Kokoro among them) as objects like { id, name }. Everything downstream — the API contract and the
+// voice <Select> — expects strings, and an object reaching the picker crashes the settings page, so
+// flatten to ids here and drop anything unrecognisable.
+const normalizeVoices = (raw: unknown): string[] => {
+  if (!Array.isArray(raw)) return [];
+  return raw
+    .map((entry) => {
+      if (typeof entry === 'string') return entry;
+      if (entry && typeof entry === 'object') {
+        const { id, name, voice_id: voiceId } = entry as Record<string, unknown>;
+        for (const candidate of [id, voiceId, name]) if (typeof candidate === 'string' && candidate) return candidate;
+      }
+      return null;
+    })
+    .filter((voice): voice is string => voice !== null);
+};
+
 ttsRouter.post('/voices', async (ctx) => {
   const body = await ctx.req.json<{ provider: string; url?: string; apiKey?: string; model?: string }>();
 
@@ -51,8 +69,8 @@ ttsRouter.post('/voices', async (ctx) => {
         headers: { 'xi-api-key': body.apiKey },
       });
       if (!res.ok) return ctx.json({ error: `ElevenLabs error: ${res.status}` }, 500);
-      const json = (await res.json()) as { voices: { voice_id: string; name: string }[] };
-      return ctx.json({ voices: json.voices.map((v) => v.voice_id) });
+      const json = (await res.json().catch(() => null)) as { voices?: unknown } | null;
+      return ctx.json({ voices: normalizeVoices(json?.voices) });
     }
 
     // OpenAI-compatible: try local server first, fallback to HuggingFace
@@ -63,8 +81,9 @@ ttsRouter.post('/voices', async (ctx) => {
     // Try /v1/audio/voices on the local server
     const localRes = await fetch(`${body.url.replace(/\/+$/, '')}/v1/audio/voices`, { headers }).catch(() => null);
     if (localRes?.ok) {
-      const json = (await localRes.json()) as { voices: string[] };
-      return ctx.json({ voices: json.voices });
+      const json = (await localRes.json().catch(() => null)) as { voices?: unknown } | null;
+      const voices = normalizeVoices(json?.voices);
+      if (voices.length > 0) return ctx.json({ voices });
     }
 
     // Fallback: get model repo from /v1/models, then list voices from HuggingFace