feat: Improve tensor handling in transcription for better model compatibility

app/services/asr.py

@@ -255,17 +255,19 @@ def transcribe(
             return_tensors="pt",
             language=language,
         )
-        inputs = {
-            k: v.to(_model.device, dtype=_model.dtype) if hasattr(v, "to") else v
-            for k, v in inputs.items()
-        }
+        # BatchFeature.to() casts only floating-point tensors to model.dtype,
+        # leaving integer tensors (masks / ids) intact. Casting every tensor
+        # to bf16 (the old behavior) corrupts those and breaks generation.
+        inputs = inputs.to(_model.device, dtype=_model.dtype)
 
         with torch.inference_mode():
             outputs = _model.generate(
                 **inputs,
                 max_new_tokens=max_new_tokens,
             )
-        text = _processor.decode(outputs[0], skip_special_tokens=True).strip()
+        # 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()
 
         return {
             "transcript": text,