Fix virtual function issue with CTC decoder

Currently, creating CTCDecoder object by passing a language model to
`lm` argument without assigning it to a variable elsewhere causes
`RuntimeError: Tried to call pure virtual function "LM::start"`.

According to discussions on PyBind11, (
pybind/pybind11#4013 and
pybind/pybind11#2839
) this is due to Python object garbage-collected by the time
it's used by code implemented in C++. It attempts to call
methods defined in Python, which overrides the base pure virtual
function, but the object which provides this override gets
deleted by garbage collrector, as the original object is not
reference counted.

This commit fixes this by simply assiging the given `lm` object
as an attribute of CTCDecoder class.

Address #3218

Author: mthrok

test/torchaudio_unittest/models/decoder/ctc_decoder_test.py

@@ -176,6 +176,7 @@ def test_lm_lifecycle(self):
         https://github.com/pytorch/audio/issues/3218
         """
         from torchaudio.models.decoder import ctc_decoder
+
         from .ctc_decoder_utils import CustomZeroLM
 
         decoder = ctc_decoder(

torchaudio/models/decoder/_ctc_decoder.py

@@ -269,6 +269,12 @@ def __init__(
             )
         else:
             self.decoder = _LexiconFreeDecoder(decoder_options, lm, silence, self.blank, transitions)
+        # https://github.com/pytorch/audio/issues/3218
+        # If lm is passed like rvalue reference, the lm object gets garbage collected,
+        # and later call to the lm fails.
+        # This ensures that lm object is not deleted as long as the decoder is alive.
+        # https://github.com/pybind/pybind11/discussions/4013
+        self.lm = lm
 
     def _get_tokens(self, idxs: torch.IntTensor) -> torch.LongTensor:
         idxs = (g[0] for g in it.groupby(idxs))