feat: Upgrade Whisper model to medium and increase beam size

- Change default model from small to medium for better accuracy
- Increase beam_size from 5 to 8 for improved transcription quality
- Add Whisper environment variables to start.sh for centralized config

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

sidecar/transcriber.py

@@ -368,7 +368,7 @@ class Transcriber:
             segments, info = self.model.transcribe(
                 audio_path,
                 language="zh",  # Use "nan" for Taiwanese/Hokkien, "zh" for Mandarin
-                beam_size=5,
+                beam_size=8,
                 vad_filter=True,
                 word_timestamps=add_punctuation,
                 # Anti-hallucination settings

start.sh

@@ -22,6 +22,11 @@ SIDECAR_DIR="$PROJECT_DIR/sidecar"
 # Port 設定
 BACKEND_PORT=8000
 
+# Whisper 語音轉文字設定
+export WHISPER_MODEL="medium"      # 模型大小: tiny, base, small, medium, large
+export WHISPER_DEVICE="cpu"        # 執行裝置: cpu, cuda
+export WHISPER_COMPUTE="int8"      # 運算精度: int8, float16, float32
+
 # PID 檔案
 PID_FILE="$PROJECT_DIR/.running_pids"