Update app.py

app.py

@@ -78,17 +78,21 @@ def transcribe_audio(audio):
     )
     inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
 
+    MAX_DECODER_TOKENS = 448
+    START_TOKENS = 4
+    max_new_tokens = MAX_DECODER_TOKENS - START_TOKENS  # 444
+    
     with torch.no_grad():
-        # Force Yoruba transcription
         generated_ids = model.generate(
             **inputs,
             task="transcribe",
-            language="yo",          # Yoruba ISO-639-1 code
-            max_new_tokens=512,
+            language="yo",
+            max_new_tokens=max_new_tokens,
             temperature=0.0,
             no_repeat_ngram_size=3
         )
 
+
     text = processor.batch_decode(
         generated_ids,
         skip_special_tokens=True