update

app.py

@@ -7,12 +7,14 @@ import requests
 import subprocess
 from pathlib import Path
 import torchaudio
+import torch
+import pyloudnorm as pyln
 
-from model import Jamify
+from model import Jamify, normalize_audio
 from utils import json_to_text, text_to_json, convert_text_time_to_beats, convert_text_beats_to_time, convert_text_beats_to_time_with_regrouping, text_to_words, beats_to_text_with_regrouping, round_to_quarter_beats
 
 def crop_audio_to_30_seconds(audio_path):
-    """Crop audio to first 30 seconds and return path to temporary cropped file"""
+    """Crop audio to first 30 seconds, normalize, and return path to temporary cropped file"""
     if not audio_path or not os.path.exists(audio_path):
         return None
     
@@ -29,11 +31,20 @@ def crop_audio_to_30_seconds(audio_path):
         else:
             cropped_waveform = waveform
         
+        # Resample to 44100 Hz if needed (to match prediction pipeline)
+        if sample_rate != 44100:
+            resampler = torchaudio.transforms.Resample(sample_rate, 44100)
+            cropped_waveform = resampler(cropped_waveform)
+            sample_rate = 44100
+        
+        # Apply the same normalization as the prediction pipeline
+        normalized_waveform = normalize_audio(cropped_waveform)
+        
         # Save to temporary file
         with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as temp_file:
             temp_path = temp_file.name
         
-        torchaudio.save(temp_path, cropped_waveform, sample_rate)
+        torchaudio.save(temp_path, normalized_waveform, sample_rate)
         return temp_path
         
     except Exception as e:
@@ -196,7 +207,18 @@ default_audio_display = crop_audio_to_30_seconds(default_audio) if default_audio
 # Gradio interface
 with gr.Blocks() as demo:
     gr.Markdown("# Jamify: Music Generation from Lyrics and Style")
-    gr.Markdown("Provide your lyrics, a style reference (either an audio file or a text prompt), and a desired duration to generate a song.")
+    gr.Markdown("Provide your lyrics, an audio style reference, and a desired duration to generate a song.")
+    
+    # Helpful reminder for users
+    gr.Markdown("""
+    💡 **Demo Tip**: Don't start from scratch! Use the sample examples below as templates:
+    - Click any sample to load its lyrics and audio style
+    - **Edit the lyrics**: Change words, modify timing, or adjust the structure
+    - **Experiment with timing**: Try different word durations or beats
+    - **Mix and match**: Use lyrics from one example with audio style from another
+    
+    This approach is much easier than creating everything from zero!
+    """)
     
     # State to track selected example (-1 means "Make Your Own" is selected, 0 is first example)
     selected_example = gr.State(0 if examples else -1)
@@ -211,8 +233,8 @@ with gr.Blocks() as demo:
         with gr.Row():
             example_buttons = []
             for i, example in enumerate(examples):
-                # Use consistent button width and truncate long IDs if needed
-                button_text = example['id'][:12] + "..." if len(example['id']) > 15 else example['id']
+                # Use consistent button width with 10 character limit
+                button_text = example['id'][:10] if len(example['id']) <= 10 else example['id'][:9] + "…"
                 # First button starts as primary (selected), others as secondary
                 initial_variant = "primary" if i == 0 else "secondary"
                 button = gr.Button(

jam_infer.yaml

@@ -23,10 +23,10 @@ evaluation:
     cfg_range:
       - 0.05
       - 1
-    fix_dual_cfg: true
+    # fix_dual_cfg: true
     dual_cfg:
       - 4.7
-      - 2.6
+      - 2.5
     steps: 50
 
 model: