diff --git a/install.sh b/install.sh
index 07be64269..ca7de7387 100755
--- a/install.sh
+++ b/install.sh
@@ -322,7 +322,7 @@ step_polkit_rules() {
 }
 
 step_voice_install() {
-  apt-get install -y -qq espeak-ng libsndfile1
+  apt-get install -y -qq espeak-ng libsndfile1 cmake build-essential
   bash "$PROJECT_DIR/scripts/install-voice.sh"
   echo "  Voice pipeline installed"
   bash "$PROJECT_DIR/scripts/setup-optimizations.sh"
diff --git a/scripts/install-voice.sh b/scripts/install-voice.sh
index dd7e331d0..64387aa4d 100755
--- a/scripts/install-voice.sh
+++ b/scripts/install-voice.sh
@@ -89,11 +89,21 @@ fi
 # ── Step 4: Install Kokoro TTS ───────────────────────────────────────────────
 
 echo "[4/7] Installing Kokoro TTS..."
-su - "$CLAWBOX_USER" -c "$PIP install --user 'numpy<2' 'transformers<5' kokoro soundfile Pillow" 2>&1 | tail -3
+# Install kokoro first, then force transformers<5 as a separate step.
+# pip 22's resolver won't downgrade huggingface-hub (pulled in by faster-whisper)
+# to satisfy transformers<5 in a single command, so it silently picks transformers 5.x.
+su - "$CLAWBOX_USER" -c "$PIP install --user 'numpy<2' kokoro soundfile 'Pillow>=10'" 2>&1 | tail -3
+su - "$CLAWBOX_USER" -c "$PIP install --user 'transformers<5'" 2>&1 | tail -3
 
 # ── Step 5: Pre-download models ─────────────────────────────────────────────
 
 echo "[5/7] Pre-downloading Whisper model (base)..."
+# Clear corrupted cache (0-byte blobs from failed/rate-limited HF downloads)
+WHISPER_CACHE="$CLAWBOX_HOME/.cache/huggingface/hub/models--Systran--faster-whisper-base"
+if [ -d "$WHISPER_CACHE/blobs" ] && find "$WHISPER_CACHE/blobs" -maxdepth 1 -type f -empty | grep -q .; then
+  echo "  Clearing corrupted Whisper model cache..."
+  rm -rf "$WHISPER_CACHE"
+fi
 DEVICE="cpu"
 COMPUTE="auto"
 if $HAS_CUDA; then