feat: add GPU support for transcription and improve error handling

- Introduced a new function `transcribe_one_with_gpu` to enable GPU acceleration for audio transcription.
- Enhanced error handling during Whisper model loading to provide clearer messages and support for unsupported platforms.
- Updated the `make_prompt` function to utilize the new GPU-enabled transcription function.

apps/audio_cloning/vallex/main.py

@@ -170,17 +170,20 @@ if not os.path.exists(OUTPUT_DIR_WHISPER):
 try:
     logger.info("Loading Whisper model...")
     model_name = "tiny"
-    whisper_model = whisper.load_model(model_name, download_root=OUTPUT_DIR_WHISPER).to(
-        device
-    )
+    whisper_model = whisper.load_model(model_name, download_root=OUTPUT_DIR_WHISPER)
     logger.info("Whisper model loaded successfully")
+except NotImplementedError as e:
+    logger.error("Error on loading Whisper model: %s", e)
+    raise Exception(
+        f"Whisper model {model_name} is not supported on this platform."
+    ) from e
 except Exception as e:
     logger.error("Error on loading Whisper model: %s", e)
     raise Exception(
         "\n Whisper download failed or damaged, please go to "
-        "'https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt'"
-        "\n manually download model and put it to {} .".format(os.getcwd() + "/whisper")
-    )
+        f"'https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/{model_name}.pt'"
+        f"\n manually download model and put it to {OUTPUT_DIR_WHISPER}."
+    ) from e
 
 # Initialize Voice Presets
 logger.info("Initializing Voice Presets...")
@@ -235,6 +238,50 @@ def transcribe_one(model, audio_path):
     return lang, text_pr
 
 
+@spaces.GPU(duration=120)
+def transcribe_one_with_gpu(model, audio_path):
+    model.eval()
+
+    # ZeroGPU では GPU 初期化/移動は関数内で
+    if torch.cuda.is_available():
+        model = model.to("cuda", non_blocking=True)
+        use_fp16 = True
+        dev = torch.device("cuda")
+    else:
+        use_fp16 = False
+        dev = torch.device("cpu")
+
+    # 推論は grad 無効に（速くて軽い）
+    with torch.inference_mode():
+        # 30 秒にパディング/トリム
+        audio = whisper.load_audio(audio_path)
+        audio = whisper.pad_or_trim(audio)
+
+        # ログメルを作成（最初は CPU の密テンソル想定）
+        mel = whisper.log_mel_spectrogram(audio)
+        mel = mel.to(dev, non_blocking=True)
+
+        # 言語推定
+        _, probs = model.detect_language(mel)
+        lang = max(probs, key=probs.get)
+        print(f"Detected language: {lang}")
+
+        # デコード
+        options = whisper.DecodingOptions(
+            temperature=1.0,
+            best_of=5,
+            fp16=use_fp16,
+            sample_len=150,
+        )
+        result = whisper.decode(model, mel, options)
+
+    text_pr = result.text
+    if text_pr.strip(" ")[-1] not in "?!.,。，？！。、":
+        text_pr += "."
+    return lang, text_pr
+
+
+@spaces.GPU(duration=120)
 def make_npz_prompt(name, uploaded_audio, recorded_audio, transcript_content):
     global model, text_collater, text_tokenizer, audio_tokenizer
     clear_prompts()
@@ -290,7 +337,7 @@ def make_prompt(name, wav, sr, save=True):
         wav = wav.unsqueeze(0)
     assert wav.ndim and wav.size(0) == 1
     torchaudio.save(f"./prompts/{name}.wav", wav, sr)
-    lang, text = transcribe_one(whisper_model, f"./prompts/{name}.wav")
+    lang, text = transcribe_one_with_gpu(whisper_model, f"./prompts/{name}.wav")
     lang_token = lang2token[lang]
     text = lang_token + text + lang_token
     with open(f"./prompts/{name}.txt", "w", encoding="utf-8") as f: