fix: improve audio compatibility, add unique filenames, and add download button in web app

main.py

@@ -0,0 +1,6 @@
+def main():
+    print("Hello from cosyvoice3-talkingflowerzh!")
+
+
+if __name__ == "__main__":
+    main()

test_debug.py

@@ -0,0 +1,12 @@
+import sys
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "2"
+sys.path.insert(0, "/home/mason/Developer/Projects/WonderFlower/CosyVoice3-TalkingFlowerZH")
+sys.path.insert(0, "/home/mason/Developer/Projects/WonderFlower/CosyVoice3-TalkingFlowerZH/third_party/Matcha-TTS")
+
+import importlib.util
+spec = importlib.util.spec_from_file_location("app", "/home/mason/Developer/Projects/WonderFlower/CosyVoice3-TalkingFlowerZH/web-app/app.py")
+app = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(app)
+
+print(app.generate_audio("你好呀！我是会说话的花朵，很高兴认识你！", 42, 1.0))

web-app/.gitignore

@@ -0,0 +1 @@
+outputs/

web-app/app.py

@@ -1,8 +1,10 @@
 import os
 import sys
 import time
+import uuid
 import random
 import torch
+import numpy as np
 import soundfile as sf
 import gradio as gr
 
@@ -12,6 +14,10 @@ _PARENT = os.path.dirname(_HERE)
 sys.path.insert(0, _PARENT)
 sys.path.insert(0, os.path.join(_PARENT, "third_party/Matcha-TTS"))
 
+# Create an output directory for generated audio
+OUTPUT_DIR = os.path.join(_HERE, "outputs")
+os.makedirs(OUTPUT_DIR, exist_ok=True)
+
 import onnxruntime
 import transformers
 
@@ -53,6 +59,7 @@ model = None
 def load_model():
     global model
     if model is None:
+        print(f"Loading model from {MODEL_DIR}...")
         model = CosyVoice3(MODEL_DIR)
     return model
 
@@ -81,28 +88,50 @@ def remove_tail_click(audio, sr, search_s=0.20, burst_thresh=0.05,
 
 def generate_audio(text, seed, speed):
     if not text:
-        return None, "Please enter some text."
+        return None, None, "Please enter some text."
     
     set_all_random_seed(seed)
     
     try:
-        model = load_model()
+        model_instance = load_model()
+        
+        # Determine if we should use frontend based on the presence of special tokens
+        # Note: cosyvoice internal logic also handles this, but we mirror it here for clarity
+        use_frontend = not ('<|' in text and '|>' in text)
         
-        for output in model.inference_sft(
-            INSTRUCT + text,
+        full_text = INSTRUCT + text
+        
+        for output in model_instance.inference_sft(
+            full_text,
             spk_id=SPK_ID,
             stream=False,
             speed=speed,
-            text_frontend=False,
+            text_frontend=False, # We pass False here because we already prepended INSTRUCT
         ):
-            audio = remove_tail_click(output["tts_speech"], model.sample_rate)
+            audio = remove_tail_click(output["tts_speech"], model_instance.sample_rate)
+            
+            # Convert to numpy and normalize
+            audio_np = audio.squeeze(0).cpu().numpy()
+            
+            # Simple normalization to avoid clipping and improve compatibility
+            max_val = np.abs(audio_np).max()
+            if max_val > 1.0:
+                audio_np = audio_np / max_val
+            
+            # Convert to 16-bit PCM for widest compatibility
+            audio_int16 = (audio_np * 32767).astype(np.int16)
+            
+            # Unique filename for each request
+            request_id = str(uuid.uuid4())[:8]
+            output_path = os.path.join(OUTPUT_DIR, f"output_{request_id}.wav")
             
-            output_path = os.path.join(_HERE, "output.wav")
-            sf.write(output_path, audio.squeeze(0).cpu().numpy(), model.sample_rate)
-            return output_path, f"Success! (Seed: {seed})"
+            sf.write(output_path, audio_int16, model_instance.sample_rate)
+            return output_path, output_path, f"Success! (Seed: {seed})"
             
     except Exception as e:
-        return None, f"Error: {str(e)}"
+        import traceback
+        traceback.print_exc()
+        return None, None, f"Error: {str(e)}"
 
 # Gradio UI Theme & Setup (inspired by Talking-Flower)
 custom_css = """
@@ -144,14 +173,15 @@ with gr.Blocks(title="Talking Flower TTS", css=custom_css) as demo:
                 audio_output = gr.Audio(
                     label="输出音频 (Generated Audio)",
                     type="filepath",
-                    interactive=False, # Enables native download button
+                    interactive=False,
                     elem_classes="wonder-card"
                 )
+                download_output = gr.File(label="下载文件 (Download Wav)", elem_classes="wonder-card")
                 status_output = gr.Textbox(label="Status", interactive=False, elem_classes="wonder-card")
                 
         gr.Markdown("---")
         gr.Markdown("### 🧠 Model Architecture (CosyVoice 3 Sub-Models)")
-        gr.HTML("""
+        gr.HTML(\"\"\"
         <div class="model-arch">
             <p>This repository uses a three-stage cascade architecture for zero-shot and supervised text-to-speech:</p>
             <ol>
@@ -160,12 +190,12 @@ with gr.Blocks(title="Talking Flower TTS", css=custom_css) as demo:
                 <li><strong>HIFT (HiFi-GAN Vocoder) <code>hift.pt</code></strong>: A high-fidelity generative adversarial network that converts the Mel-spectrograms into the final raw audio waveform.</li>
             </ol>
         </div>
-        """)
+        \"\"\")
 
         generate_btn.click(
             fn=generate_audio,
             inputs=[text_input, seed_input, speed_input],
-            outputs=[audio_output, status_output]
+            outputs=[audio_output, download_output, status_output]
         )
 
 if __name__ == "__main__":