diff --git a/client/lucy-desktop/src/main/index.ts b/client/lucy-desktop/src/main/index.ts
index 3f73c8b..e892e47 100644
--- a/client/lucy-desktop/src/main/index.ts
+++ b/client/lucy-desktop/src/main/index.ts
@@ -95,7 +95,14 @@ async function ensureTtsServer() {
   } catch { /* selbst starten */ }
   console.log("[lucy] starte pocket_server:", PY, "(cwd:", TTS_DIR + ")")
   ttsProc = spawn(PY, ["-m", "uvicorn", "pocket_server:app", "--host", "127.0.0.1", "--port", PORT],
-    { cwd: TTS_DIR, windowsHide: true, env: { ...process.env } })
+    { cwd: TTS_DIR, windowsHide: true, env: {
+      ...process.env,
+      // Bench-Optimum aus Phase B2 (docs/LUCY_TTS_PLAN.md): 4 Worker x 2 Threads = RTF 0,44
+      // statt 0,67 seriell. Ohne diese Envs startete der Server SERIELL (workers=0) ->
+      // TTS-TTFB 1,7-2,3 s je Antwort (Realtest 02.07.). Per Umgebung überschreibbar.
+      LUCY_WORKERS: process.env.LUCY_WORKERS || "4",
+      LUCY_WORKER_THREADS: process.env.LUCY_WORKER_THREADS || "2",
+    } })
   ttsProc.stdout?.on("data", (d) => console.log("[tts]", d.toString().trimEnd()))
   ttsProc.stderr?.on("data", (d) => console.log("[tts]", d.toString().trimEnd()))
   ttsProc.on("error", (e) => console.error("[lucy] pocket_server-Spawn fehlgeschlagen:", e))