diff --git a/Dockerfile b/Dockerfile
index d3fc9e5..62fe8f4 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -16,12 +16,14 @@ COPY pyproject.toml .
 COPY src/ src/
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install .
-# Voice dependencies (faster-whisper, Kokoro TTS, soundfile) — activated at runtime via VOICE_ENABLED
-# Install CPU-only torch first so pip doesn't pull full CUDA wheels (~2 GB) for kokoro/transformers.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install torch --index-url https://download.pytorch.org/whl/cpu \
-    && pip install faster-whisper kokoro soundfile \
-    && python -m spacy download en_core_web_sm
+
+# Voice deps (faster-whisper, kokoro, torch, soundfile, spacy) are NOT
+# installed in the production image. They added ~2 GB and pulled numpy<2,
+# which has no cp314 wheel and broke the slim-image build.
+# Voice paths in code are lazily guarded; with VOICE_ENABLED=false (default),
+# the app starts cleanly. With voice enabled, missing imports log and
+# the feature gracefully degrades. To re-enable voice in a build, install
+# the `voice` extra: `pip install .[voice]` — and add an extras step here.
 
 # Build the fable-mcp wheel so it can be served for download
 COPY fable-mcp/ fable-mcp/