diff --git a/Dockerfile b/Dockerfile
index 62fe8f4..84b5971 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -17,13 +17,14 @@ COPY src/ src/
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install .
 
-# 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.
+# Speech-to-text (faster-whisper + soundfile). faster-whisper is pure
+# Python; the actual inference engine is ctranslate2 (C++) which has
+# cp314 wheels as of v4.7.2 (2026-05-19). No torch needed — ctranslate2
+# does its own CPU inference. Image add: ~150 MB.
+# TTS (piper-tts) is installed in a separate later step so STT can
+# stand alone if TTS becomes problematic.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install faster-whisper soundfile
 
 # Build the fable-mcp wheel so it can be served for download
 COPY fable-mcp/ fable-mcp/