FST-based forced alignment

This section describes how to perform FST-based forced alignment with models trained by CTC loss.

We use CTC FORCED ALIGNMENT API TUTORIAL from torchaudio as a reference in this section.

Different from torchaudio, we use an FST-based approach.

Contents:

• Two approaches
  • Differences between the two approaches
• Kaldi-based forced alignment
  • Prepare the environment
  • Get the test data
  • Compute log_probs
  • Create token2id and id2token
  • Create word2id and id2word
  • Generate lexicon-related files
  • Convert transcript to an FST graph
  • Force aligner
  • Segment each word using the computed alignments
  • Summary
• k2-based forced alignment