diff --git a/homelab/docker-compose.yml b/homelab/docker-compose.yml
index 878c69c..c1dfded 100644
--- a/homelab/docker-compose.yml
+++ b/homelab/docker-compose.yml
@@ -414,21 +414,27 @@ services:
 
   # ── pocket-tts (CPU TTS) ────────────────────────────────────────────────────
   # Lightweight CPU-only TTS using kyutai-labs/pocket-tts.
-  # OpenAI-compatible: POST /v1/audio/speech on port 8000.
-  # Voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma.
+  # Image is built locally on homelab from https://github.com/kyutai-labs/pocket-tts
+  # (no prebuilt image published): cd /tmp && git clone --depth=1 https://github.com/kyutai-labs/pocket-tts.git && docker build -t pocket-tts:latest /tmp/pocket-tts
+  # OpenAI-compatible: POST /tts (multipart form) on port 8000.
+  # Voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma, etc.
   # Not currently used by the runner (runner uses kokoro-fastapi), but available
   # for experimentation / fallback.
   pocket-tts:
-    image: ghcr.io/kyutai-labs/pocket-tts:latest
+    image: pocket-tts:latest
     restart: unless-stopped
+    command: ["uv", "run", "pocket-tts", "serve", "--host", "0.0.0.0"]
     expose:
       - "8000"
+    volumes:
+      - pocket_tts_cache:/root/.cache/pocket_tts
+      - hf_cache:/root/.cache/huggingface
     healthcheck:
       test: ["CMD", "curl", "-sf", "http://localhost:8000/health"]
       interval: 30s
       timeout: 10s
       retries: 5
-      start_period: 60s
+      start_period: 120s
 
   # ── Watchtower ──────────────────────────────────────────────────────────────
   # Auto-updates runner image when CI pushes a new tag.
@@ -453,3 +459,5 @@ volumes:
   prometheus_data:
   loki_data:
   grafana_data:
+  pocket_tts_cache:
+  hf_cache: