Update app.py

app.py

@@ -1,46 +1,52 @@
 import gradio as gr
 from TTS.utils.synthesizer import Synthesizer
+from TTS.tts.models.vits import Vits
+from TTS.tts.configs.vits_config import VitsConfig
 from huggingface_hub import hf_hub_download
+import torch
 import os
 import json
-import gc
 
 def load_eng_model():
     repo_id = "E-motionAssistant/text-to-speech-VITS-english"
     print(f"--- Downloading English Model ---")
     
-    # 1. Download original files
     model_path = hf_hub_download(repo_id=repo_id, filename="best_model.pth")
     config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
 
-    # 2. Read the config file and manually change the number to 137
-    with open(config_path, 'r') as f:
-        config_dict = json.load(f)
+    # 1. Load and fix config
+    config = VitsConfig()
+    config.load_json(config_path)
+    config.model_args.num_chars = 137 # The magic number
     
-    print("Fixing character count: 131 -> 137")
-    config_dict["model_args"]["num_chars"] = 137
+    # 2. Build the Model Architecture Manually
+    print("Building model architecture...")
+    model = Vits.init_from_config(config)
     
-    # 3. Save this as a NEW file (a physical path the library will accept)
-    fixed_config_path = os.path.join(os.getcwd(), "fixed_config.json")
-    with open(fixed_config_path, 'w') as f:
-        json.dump(config_dict, f)
+    # 3. Load the checkpoint weights
+    print("Loading weights (using non-strict mode to bypass mismatch)...")
+    checkpoint = torch.load(model_path, map_location="cpu")
+    
+    # This is the line that solves the "Size Mismatch" crash
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.eval()
 
-    # 4. Initialize the Synthesizer using the PATH to our new file
-    print("Building Synthesizer with fixed_config.json path...")
+    # 4. Wrap it in a Synthesizer
     syn = Synthesizer(
         tts_checkpoint=model_path,
-        tts_config_path=fixed_config_path, # This is now a STRING path
+        tts_config_path=config_path,
         use_cuda=False
     )
+    # Overwrite the internal model with our forced one
+    syn.tts_model = model
     
-    gc.collect()
     return syn
 
-# --- Global Initialization ---
+# --- Initialization ---
 print("Starting English TTS Startup...")
 try:
     eng_tts = load_eng_model()
-    print("--- SUCCESS: ENGLISH MODEL LOADED ---")
+    print("--- SUCCESS: SYSTEM READY ---")
 except Exception as e:
     print(f"CRITICAL ERROR: {e}")
     eng_tts = None
@@ -50,9 +56,7 @@ def generate_voice(text):
         return None
     try:
         output_path = os.path.join(os.getcwd(), "output.wav")
-        # Generate the audio samples
         wav = eng_tts.tts(text=str(text))
-        # Save samples to a .wav file
         eng_tts.save_wav(wav, output_path)
         return output_path
     except Exception as e:
@@ -62,9 +66,9 @@ def generate_voice(text):
 # Gradio Interface
 demo = gr.Interface(
     fn=generate_voice,
-    inputs=gr.Textbox(label="Input English Text", placeholder="Hello world!"),
-    outputs=gr.Audio(label="Synthesized Speech", type="filepath"),
-    title="English VITS TTS"
+    inputs=gr.Textbox(label="Input English Text"),
+    outputs=gr.Audio(label="Audio Output", type="filepath"),
+    title="English VITS TTS (Forced Load)"
 )
 
 if __name__ == "__main__":