igbo asr fix

app.py

@@ -106,12 +106,11 @@ def _get_igbo_asr():
         logger.warning("HF_TOKEN not set - Igbo ASR model requires authentication")
         return None, None
     
-    
     hf_token = hf_token.strip()
     
     try:
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        logger.info("Lazy-loading Igbo ASR model (gated model)...")
+        logger.info("Lazy-loading Igbo ASR model...")
         igbo_processor = WhisperProcessor.from_pretrained("NCAIR1/Igbo-ASR", token=hf_token)
         igbo_model = WhisperForConditionalGeneration.from_pretrained("NCAIR1/Igbo-ASR", token=hf_token)
         igbo_model.to(device)
@@ -132,13 +131,12 @@ def _run_whisper(model: WhisperForConditionalGeneration, proc: WhisperProcessor,
         generation_kwargs = {
             "max_length": 448,
             "num_beams": 1,
-            "do_sample": False,
-            "early_stopping": True
+            "do_sample": False
         }
         
-        
         if language == "igbo" or "igbo" in str(model.config).lower():
-            generation_kwargs["language"] = "igbo"
+            pass
+        else:
             generation_kwargs["task"] = "transcribe"
         
         with torch.no_grad():
@@ -194,7 +192,6 @@ def preprocess_audio_ffmpeg(audio_data: bytes, target_sr: int = 16000) -> np.nda
 def speech_to_text(audio_data: bytes) -> str:
     audio_array = preprocess_audio_ffmpeg(audio_data)
     
-    # Try Igbo ASR first for better Igbo detection
     igbo_result = _get_igbo_asr()
     if igbo_result[0] is not None and igbo_result[1] is not None:
         igbo_model, igbo_proc = igbo_result
@@ -203,9 +200,8 @@ def speech_to_text(audio_data: bytes) -> str:
             logger.info("Using Igbo ASR result")
             return igbo_text
     
-    # Fallback to MMS for other languages
     mms_result = _get_mms()
-    if mms_result[0] is not None and mms_result[1] is not None:
+    if mms_result and mms_result[0] is not None and mms_result[1] is not None:
         mms_model, mms_proc = mms_result
         mms_text = _run_mms(mms_model, mms_proc, audio_array)
         if mms_text and mms_text.strip():