The Nemotron-3.5-ASR-Streaming-0.6B model has poor Chinese speech recognition performance.

[liujixingit]

I called the official test demo for speech recognition testing。as follow:
python examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py model_path=/root/.cache/modelscope/hub/models/nv-community/nemotron-3.5-asr-streaming-0.6b/nemotron-3.5-asr-streaming-0.6b.nemo audio_file=/opt/ai/liujixin/ASR/case5_sohu.wav batch_size=1 target_lang=auto att_context_size="[56,13]" strip_lang_tags=true compare_vs_offline=true amp=true debug_mode=true

audio file.
case5_sohu.wav

ASR output result：

the ground truth as follow:

I'm not sure if it's a problem with my invocation or a performance flaw in the model itself.

[rtxlovelace-collab]

How did you get around this error?

Traceback (most recent call last):
  File "/home/ubuntu/nvidia-asr/NeMo/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py", line 117, in <module>
    from nemo.collections.asr.parts.utils.streaming_utils import CacheAwareStreamingAudioBuffer
  File "/home/ubuntu/nvidia-asr/NeMo/nemo/collections/asr/parts/utils/streaming_utils.py", line 2467
    shifted_indices = shifted_indices[..., *[None for _ in range(len(self.dim_shape))]]
                                           ^

[liujixingit]

@rtxlovelace-collab I did not encounter the program execution error you mentioned.

[liujixingit]

With respect to the issue I raised above, I believe NVIDIA's algorithm ought not to yield such poor results. I suspect that my code may have been executed incorrectly in some respect.

[liujixingit]

I performed the execution using the officially recommended container image nvcr.io/nvidia/nemo:26.02 in conjunction with the most recent Git source code.