--- a/app.py
+++ b/app.py
@@ -206,15 +206,29 @@
                 progress(0.5, 'Using provided reference text. Encoding audio...')
             else:
                 progress(0.25, 'Transcribing reference audio with Whisper...')
+                # llasaモデルをCPUに移動してVRAMを解放 VRAMのキャッシュをクリア
+                print("[VRAM_SWAP] Llasa model moved to CPU.")
+                model.to("cpu")
+                torch.cuda.empty_cache()
                 global whisper_turbo_pipe
                 if whisper_turbo_pipe is None:
                     whisper_turbo_pipe = pipeline(
                         "automatic-speech-recognition",
-                        model="openai/whisper-large-v3-turbo",
+                        model="litagin/anime-whisper",
                         torch_dtype=torch.float16,
                         device=device.type,
+                        chunk_length_s=30.0,
+                        batch_size=64,
                     )
+                else:
+                    print("[Whisper] Moving Whisper model to CUDA (FP16) for transcription.")
+                    whisper_turbo_pipe.model.to("cuda").half()
                 prompt_text = whisper_turbo_pipe(prompt_wav[0].cpu().numpy())['text'].strip()
+                # whisperモデルをCPUに移動してVRAMを解放 llasaモデルをCUDAに戻してVRAMのキャッシュをクリア
+                print("[VRAM_SWAP] Whisper model moved to CPU. Llasa model moved to CUDA")
+                whisper_turbo_pipe.model.to("cpu").half()
+                model.to("cuda")
+                torch.cuda.empty_cache()
                 progress(0.5, 'Transcribed! Encoding reference audio...')
 
             # Encode prompt audio to codec tokens