Update app.py

app.py

@@ -287,15 +287,12 @@ word_syllables = {
     "dula": ["du", "la"]
 }
 
-import torch.nn.functional as F
-
-def transcribe(audio_file, expected_word=None):
-    if not audio_file or not expected_word:
-        return ["No audio or expected word provided.", "0.00"]
+def transcribe(audio_file, target_word):
+    if not audio_file or not target_word:
+        return "No audio or word selected.", "0.00"
 
-    target_word = expected_word.lower().strip()
+    # Load and process audio directly
     input_values = processor(audio_file, sampling_rate=SR, return_tensors="pt").input_values.to(DEVICE)
-    
     with torch.no_grad():
         logits = model(input_values).logits
 
@@ -322,24 +319,23 @@ def transcribe(audio_file, expected_word=None):
     ref_emb = get_embedding(target_sylls)
 
     if pred_emb is None or ref_emb is None:
-        cosine_sim = 0.0
-    else:
-        cosine_sim = F.cosine_similarity(pred_emb.unsqueeze(0), ref_emb.unsqueeze(0)).item()
+        return transcription, "0.00"
+
+    score = F.cosine_similarity(pred_emb, ref_emb, dim=0).item()
+    return transcription, str(round(score, 4))
 
-    return [transcription, str(round(cosine_sim, 2))]
-	
 iface = gr.Interface(
     fn=transcribe,
     inputs=[
         gr.Audio(type="filepath", label="🎤 Record or Upload Audio"),
-        gr.Textbox(label="Expected Word (GIF Title)")  # <-- textbox for app to send
+        gr.Dropdown(sorted(word_syllables.keys()), label="Select Target Word")
     ],
     outputs=[
         gr.Textbox(label="🗣️ Transcription", interactive=False),
         gr.Textbox(label="📈 Cosine Similarity", interactive=False)
     ],
     title="Cognia Wav2Vec2 Pronunciation Practice",
-    description="Upload or record audio, and get similarity score vs the expected word."
+    description="Upload or record audio, and get similarity score vs the target Filipino word."
 )
 
 iface.launch()