feat: Normalize output from processor.decode in transcribe function for consistent string handling

app/services/asr.py

@@ -265,9 +265,13 @@ def transcribe(
                 **inputs,
                 max_new_tokens=max_new_tokens,
             )
-        # Cohere's processor.decode expects the generated tensor as-is
-        # (per the model card), not a single unwrapped sequence.
-        text = _processor.decode(outputs, skip_special_tokens=True).strip()
+        # Cohere's processor.decode is given the generated tensor as-is
+        # (per the model card). For a batched [1, seq] output it returns a
+        # *list* of strings, so normalize to a single string.
+        decoded = _processor.decode(outputs, skip_special_tokens=True)
+        if isinstance(decoded, (list, tuple)):
+            decoded = decoded[0] if decoded else ""
+        text = decoded.strip()
 
         return {
             "transcript": text,