NeMo ASR: ModuleNotFoundError 'ctc_decoders' when changing decoding strategy to Beam Search in Google Colab

I am working with NVIDIA NeMo's ASR collection (QuartzNet model) in Google Colab. I am trying to evaluate the model using CTC Beam Search without a Language Model (setting alpha=0.0).

Even after installing pyctcdecode and kenlm, I encounter a ModuleNotFoundError: No module named 'ctc_decoders' when I call change_decoding_strategy.

Environment:

• OS: Google Colab (Ubuntu 22.04 LTS)

• Python: 3.12

• NeMo version: Latest (installed via pip)

Code Snippet:

import os
from omegaconf import DictConfig

# Creating a dummy file to bypass initial FileNotFoundError for kenlm_path
with open("dummy.bin", "w") as f:
    f.write("")

beam_widths = [3, 10, 20]

for width in beam_widths:
    try:
        # Resetting decoding strategy
        quartznet_model_copy.decoding = None

        beam_cfg = DictConfig({
            "strategy": "beam",
            "beam": {
                "search_type": "default",  
                "kenlm_path": os.path.abspath("dummy.bin"),
                "width": width,            
                "alpha": 0.0,              
                "beta": 0.0,
                "return_best_hypothesis": True
            }
        })

        quartznet_model_copy.change_decoding_strategy(decoding_cfg=beam_cfg)
        # evaluate_wer call...
        
    except Exception as e:
        print(f"Error: {e}")
        raise e

ModuleNotFoundError: No module named 'ctc_decoders'
...
/usr/local/lib/python3.12/dist-packages/nemo/collections/asr/modules/beam_search_decoder.py in __init__(...)
     64             from ctc_decoders import Scorer, ctc_beam_search_decoder_batch
     65         except ModuleNotFoundError:
---> 66             raise ModuleNotFoundError(
     67                 "BeamSearchDecoderWithLM requires the installation of ctc_decoders "
     68                 "from scripts/asr_language_modeling/ngram_lm/install_beamsearch_decoders.sh"

What I've tried:

1. Installing pyctcdecode via pip, but the default search type still triggers the old ctc_decoders dependency.

2. Changing search_type to pyctcdecode, but I still get the same requirement for ctc_decoders.

Question: How can I use Beam Search in NeMo on Google Colab without having to compile the legacy ctc_decoders library? Is there a way to force NeMo to use pyctcdecode as the backend without triggering this missing module error?