TDT compute timestamps option and Extra Whitespace handling for SPE

[monica-sekoyan]

What does this PR do ?

This PR includes option for getting timestamps from TDT Decoder head, including on a segment level. This option is enabled for RNNT and CTC decodings as well. And also introduces workaround for getting tokens and id_s for extra whitespaces (even if the tokenizer was not trained in this way.) This is crucial for getting alignments in the future.

Collection: asr, common

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• To get timestamps from TDT decoding head, just change the decoding config of the model in the following way:

import nemo.collections.asr.models as nemo_asr
import omegaconf

tdt_asr_model = nemo_asr.EncDecRNNTBPEModel.from_pretrained(model_name='nvidia/parakeet-tdt_ctc-1.1b')

cfg = model.cfg.decoding
with omegaconf.open_dict(cfg):
    cfg['compute_timestamps']=True

tdt_asr_model.change_decoding_strategy(decoding_cfg=cfg)

• To control timestamp level manipulate the following field of decoding config:

cfg = model.cfg.decoding
with omegaconf.open_dict(cfg):
    cfg['compute_timestamps']=True
    cfg['rnnt_timestamp_type'] = 'all' # can be "char", "word", "segment" or "all"

tdt_asr_model.change_decoding_strategy(decoding_cfg=cfg)

• To make the SPE tokenizer (already trained) encode the text not ignoring occurring extra whitespaces (which is the default behaviour), modify the tokenizer in the following way:

model.tokenizer.ignore_extra_whitespaces = False

• To train the SPE tokenizer to make it ignore extra whitespaces both during training and further encoding, use this flag in the training script:

python scripts/tokenizers/process_asr_text_tokenizer.py --tokenizer=spe --spe_remove_extra_whitespaces

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

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

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[nithinraok]

Thanks much!

[pzelasko]

Thank you, these prompt formatting tests are much better now - readable and future-proof.

[pzelasko]

I'm not sure I understand - why is this needed?

[monica-sekoyan]

This is for already trained tokenizers (which are trained to ignore extra whitespaces). by adding this parameter we can make inference to these models not to ignore them.
So, for example, if we want to encode ASR model's prediction text with a lot of spaces, we will get this tokens (and their corresponding ids) ["_text", "_with", "_", "a", "_lot", "_of", "_spaces"]. but instead we should have got ["_text", "_with", "_", "_", "_", "_", "a", "_lot", "_of", "_spaces"] which was in fact predicted by the model and can be used for alignment purposes (I believe there are other use-cases too)

[github-actions]

[🤖]: Hi @monica-sekoyan 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully

So it might be time to merge this PR or get some approvals

I'm just a bot so I'll leave it you what to do next.

//cc @pablo-garay @ko3n1g

[nithinraok]

LGTM, Thank you!