limit 4s CPU

README.md

@@ -1,5 +1,5 @@
 ---
-title: Wav2Vec2 / Wav2small
+title: Wav2small 2.0
 emoji: 🎵
 colorFrom: blue
 colorTo: pink

app.py

@@ -369,7 +369,7 @@ ax.spines['left'].set_visible(False)
 
 def process_audio(audio_filepath):
     if audio_filepath is None:
-        
+
         return fig_error, fig_error
 
     waveform, sample_rate = librosa.load(audio_filepath, sr=None)
@@ -380,20 +380,20 @@ def process_audio(audio_filepath):
     else:
         resampled_waveform_np = waveform[None, :]
 
-    x = torch.from_numpy(resampled_waveform_np).to(torch.float)
+    x = torch.from_numpy(resampled_waveform_np[:, :64000]).to(torch.float)  # only 4s for speed
 
     with torch.no_grad():
 
         logits_dawn = dawn(x).cpu().numpy()[0, :]
 
-        logits_wavlm = base(x).cpu().numpy()[0, :]        
+        logits_wavlm = base(x).cpu().numpy()[0, :]
 
         # 17K params
         logits_wav2small = wav2small(x).cpu().numpy()[0, :]
 
 
     # --- Plot 1: Wav2Vec2 vs Wav2Small Teacher Outputs ---
-    
+
     fig, ax = plt.subplots(figsize=(10, 6))
 
     left_bars_data = logits_dawn.clip(0, 1)
@@ -460,7 +460,7 @@ def process_audio(audio_filepath):
 
     fig_2, ax_2 = plt.subplots(figsize=(10, 6))
 
-    
+
     left_bars_data = logits_wavlm.clip(0, 1)
     right_bars_data = (.5 * logits_dawn + .5 * logits_wavlm).clip(0, 1)