Update backend.py

backend.py

@@ -9,6 +9,9 @@ import reportlab
 from huggingface_hub import hf_hub_download
 from transformers import pipeline
 import soundfile as sf
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+import torch
+from datasets import load_dataset
   
 
 # -----------------------------------------------------
@@ -191,14 +194,37 @@ def export_to_pdf(result: dict, filename="FinTalk_Report.pdf"):
 
 
 def generate_tts_files(result):
-    print("🎙️ Offline TTS (espnet VITS) başlatıldı...")
-    pipe = pipeline("text-to-speech", model="espnet/kan-bayashi_ljspeech_vits")
-    keys = ["moderator_intro","bullish_view","bearish_view","moderator_wrap"]
-    for key in keys:
+    try:
+        processor = SpeechT5Processor.from_pretrained("facebook/speecht5_tts")
+        model = SpeechT5ForTextToSpeech.from_pretrained("facebook/speecht5_tts")
+        vocoder = SpeechT5HifiGan.from_pretrained("facebook/speecht5_hifigan")
+        embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+        print("🎙️ Offline TTS modeli yüklendi (SpeechT5 + HiFiGAN)")
+    except Exception as e:
+        print("⚠️ Model yüklenemedi:", e)
+        return
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+    vocoder.to(device)
+
+    # default speaker embedding
+    speaker_embedding = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0).to(device)
+
+    texts = {
+        "moderator_intro": result["moderator_intro"],
+        "bullish_view": result["bullish_view"],
+        "bearish_view": result["bearish_view"],
+        "moderator_wrap": result["moderator_wrap"]
+    }
+
+    for key, text in texts.items():
         try:
             print(f"🔊 {key} ses üretiliyor...")
-            out = pipe(result[key])
-            sf.write(f"{key}.wav", out["audio"], 22050)
-            print(f"✅ {key}.wav oluşturuldu (offline VITS)")
+            inputs = processor(text=text, return_tensors="pt").to(device)
+            speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
+            filename = f"{key}.wav"
+            sf.write(filename, speech.cpu().numpy(), samplerate=16000)
+            print(f"✅ {filename} oluşturuldu (SpeechT5 offline)")
         except Exception as e:
             print(f"TTS hatası ({key}):", e)