Add pyctcdecode to high level beam search API

[titu1994]

What does this PR do ?

Adds PyCTCDecode to high level beam search API

Collection: [ASR]

Changelog

• Adds pyctcdecode support.
• Approximates the tokenized text output since the engine does not return token indices

Usage

import nemo.collections.asr as nemo_asr

filepath = "<Path to audio file>"
kenlm_path = "<Path to ARPA or BIN KenLM File built in CHAR / WORD mode; not SUBWORD mode !>"

model = nemo_asr.models.ASRModel.from_pretrained('stt_en_conformer_ctc_large')  # type: nemo_asr.models.EncDecCTCModelBPE
# Create default decoding config
model.change_decoding_strategy(None)
decoding_cfg = model.cfg.decoding
decoding_cfg.strategy = "pyctcdecode"
decoding_cfg.beam.beam_size = 128
decoding_cfg.beam.return_best_hypothesis = False
decoding_cfg.beam.beam_alpha = 1.0
decoding_cfg.beam.beam_beta = 0.0
decoding_cfg.beam.kenlm_path = kenlm_path

# pyctcdecode config modifications
decoding_cfg.beam.pyctcdecode_cfg.beam_prune_logp = -10.
decoding_cfg.beam.pyctcdecode_cfg.token_min_logp = -5.

# Prepare beam search decoding strategy
model.change_decoding_strategy(decoding_cfg)

hypothesis = model.transcribe([filepath], return_hypotheses=True)
print("Num hypothesis per sample :", len(hypothesis[0]))
print("Result type :", hypothesis[0][0].__class__.__name__)
for candidate in range(len(hypothesis[0])):
    print(f"Beam Search {candidate + 1} :", hypothesis[0][candidate].text, hypothesis[0][candidate].score)

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

[tango4j]

Is there a particular reason to import in the function? Seems like BeamSearchDecoderWithLM is imported in the function to avoid circular import.

[titu1994]

All beam search are imported directly inside of the forward in order to prevent runtime import issue if the library is not installed.

[tango4j]

I don't see anything particular needed changes at the moment, but probably I could request some feature when I add this to diarization-ASR pipeline.

[titu1994]

Sounds good, thanks !