kaldi_examples

Sample codes and data to follow up on TIAL group meeting on 2/25. WIP

References

• Official docs
• Slides from Group Meeting
• Eleanor Chodroff's tutorial
• JHU Summer Workshop
• Random blogs I found helpful:
  • http://www.programmersought.com/tag/kaldi/
  • Using a pretrained model
  • Using and extending the ASpIRE model
  • Desh Raj's blog posts
  • ASR lecture notes from University of Edinburgh

Common Setup/Quick Start

• Install Kaldi
• Create a directory for data and experiments
• Copy cmd.sh and path.sh to this directory; this should be the same in any kaldi/egs recipes
• Example audio files are in sample_data; the audio has to be in .wav formatand mono-channel. For an example of how to do this, see convert_trim.sh

Feature Extraction

• Make sure the configurations are correct in /conf. For example:
  • I changed fbank.conf to also extract total energy by setting --use-energy=true
  • I used mfcc_hires.conf instead of the default mfcc.conf; according to the ASpiRE-related models, this gave better results in speech recognition
• comp_mfcc.sh, comp_fbank_energy.sh, and comp_pitch.sh are sample scripts to extract mfcc, fbank, and pitch features. Note that the split channel part is repeated, comment this out if you've already done this.

Decoding Using a Pretrained Model

For my purposes (conversational speech, Switchboard), I'm using the chain model trained on Fisher data. The paper that this model is based on is this one, and this blog post has some nice and detailed derivations.

Steps:

• Download a model from kaldi models. For this example, I'm using the ASpIRE chain model, version with the precompiled HCLG
• Untar it: tar xvf 0001_aspire_chain_model_with_hclg.tar.bz2
• To this recipe, copy cmd.sh and path.sh if you haven't done so
• Link common modules we'll be using (or copy all these here if you'll be editing the scripts in these directory):

  export KALDI_ROOT=/homes/ttmt001/kaldi
  ln -s $KALDI_ROOT/egs/aspire/s5/steps .
  ln -s $KALDI_ROOT/egs/aspire/s5/utils .
  ln -s $KALDI_ROOT/egs/aspire/s5/conf .
  ln -s $KALDI_ROOT/egs/aspire/s5/local .
  

• Most instructions and comments are in decode_audio.sh
  • Steps 1 and 2 can be done first and then reused (model preparation, graph compilation)
  • Steps 3 and 4 can be run after creating wav.scp and utt2spk files in my_data folder. An example of creating those is in my_data