diff --git a/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py b/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py
index 7726c2b27..981c099ef 100644
--- a/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py
+++ b/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py
@@ -120,9 +120,19 @@ def calc_drop_extra_pre_encoded(asr_model, step_num, pad_and_drop_preencoded):
 
 
 def perform_streaming(
-    asr_model, streaming_buffer, compare_vs_offline=False, debug_mode=False, pad_and_drop_preencoded=False
+    asr_model,
+    streaming_buffer,
+    compare_vs_offline=False,
+    debug_mode=False,
+    pad_and_drop_preencoded=False,
+    target_lang=None,
 ):
     batch_size = len(streaming_buffer.streams_length)
+
+    lang_ids = None
+    if target_lang is not None:
+        lang_ids = [target_lang] * batch_size
+
     if compare_vs_offline:
         # would pass the whole audio at once through the model like offline mode in order to compare the results with the stremaing mode
         # the output of the model in the offline and streaming mode should be exactly the same
@@ -141,6 +151,7 @@ def perform_streaming(
                         processed_signal=processed_signal,
                         processed_signal_length=processed_signal_length,
                         return_transcription=True,
+                        lang_ids=lang_ids,
                     )
         final_offline_tran = extract_transcriptions(transcribed_texts)
         logging.info(f" Final offline transcriptions:   {final_offline_tran}")
@@ -181,6 +192,7 @@ def perform_streaming(
                             asr_model, step_num, pad_and_drop_preencoded
                         ),
                         return_transcription=True,
+                        lang_ids=lang_ids,
                     )
 
         if debug_mode:
@@ -276,6 +288,13 @@ def main():
         help="Selects the decoder for Hybrid ASR models which has both the CTC and RNNT decoder. Supported decoders are ['ctc', 'rnnt']",
     )
 
+    parser.add_argument(
+        "--target_lang",
+        type=str,
+        default=None,
+        help="Language ID for multilingual models, e.g. pa",
+    )
+
     parser.add_argument(
         "--att_context_size",
         type=str,
@@ -376,7 +395,9 @@ def main():
             asr_model=asr_model,
             streaming_buffer=streaming_buffer,
             compare_vs_offline=args.compare_vs_offline,
+            debug_mode=args.debug_mode,
             pad_and_drop_preencoded=args.pad_and_drop_preencoded,
+            target_lang=args.target_lang,
         )
     else:
         # stream audio files in a manifest file in batched mode
@@ -409,6 +430,7 @@ def main():
                     compare_vs_offline=args.compare_vs_offline,
                     debug_mode=args.debug_mode,
                     pad_and_drop_preencoded=args.pad_and_drop_preencoded,
+                    target_lang=args.target_lang,
                 )
                 all_streaming_tran.extend(streaming_tran)
                 if args.compare_vs_offline:
diff --git a/nemo/collections/asr/models/hybrid_rnnt_ctc_models.py b/nemo/collections/asr/models/hybrid_rnnt_ctc_models.py
index 3e9f21e04..29c97e02a 100644
--- a/nemo/collections/asr/models/hybrid_rnnt_ctc_models.py
+++ b/nemo/collections/asr/models/hybrid_rnnt_ctc_models.py
@@ -163,12 +163,34 @@ class EncDecHybridRNNTCTCModel(EncDecRNNTModel, ASRBPEMixin, InterCTCMixin):
             self.ctc_decoder.unfreeze()
 
     def _transcribe_forward(self, batch: Any, trcfg: TranscribeConfig):
-        
+
+        # AI4Bharat multilingual GPU transcribe fix:
+        # transcribe() dataloader yields CPU audio tensors even when model is CUDA.
+        device = next(self.parameters()).device
+
+        if isinstance(batch, tuple):
+            batch = (
+                batch[0].to(device, non_blocking=True),
+                batch[1].to(device, non_blocking=True),
+                *batch[2:],
+            )
+        elif isinstance(batch, list):
+            batch = list(batch)
+            batch[0] = batch[0].to(device, non_blocking=True)
+            batch[1] = batch[1].to(device, non_blocking=True)
+        else:
+            raise TypeError(f"Unexpected transcribe batch type: {type(batch)}")
+
         if self.cur_decoder == "rnnt":
             return super()._transcribe_forward(batch, trcfg)
 
         # CTC Path
-        encoded, encoded_len = self.forward(input_signal=batch[0], input_signal_length=batch[1])
+        input_signal = batch[0]
+        input_signal_length = batch[1]
+        encoded, encoded_len = self.forward(
+            input_signal=input_signal,
+            input_signal_length=input_signal_length,
+        )
         if "multisoftmax" not in self.cfg.decoder:
             language_ids = None
         else:
diff --git a/nemo/collections/asr/parts/mixins/mixins.py b/nemo/collections/asr/parts/mixins/mixins.py
index 57dc95da8..9256e5444 100644
--- a/nemo/collections/asr/parts/mixins/mixins.py
+++ b/nemo/collections/asr/parts/mixins/mixins.py
@@ -584,6 +584,7 @@ class ASRModuleMixin(ASRAdapterModelMixin):
         drop_extra_pre_encoded: int = None,
         return_transcription: bool = True,
         return_log_probs: bool = False,
+        lang_ids: List[str] = None,  # AI4Bharat multilingual streaming
     ):
         """
         It simulates a forward step with caching for streaming purposes.
@@ -689,6 +690,7 @@ class ASRModuleMixin(ASRAdapterModelMixin):
                 encoded_lengths=encoded_len,
                 return_hypotheses=True,
                 partial_hypotheses=previous_hypotheses,
+                lang_ids=lang_ids,
             )
             greedy_predictions = [hyp.y_sequence for hyp in best_hyp]