Skip to content

Scripts for training acoustic and language models using CU Kids' Speech Corpus. Also includes the data preparation scripts.

Notifications You must be signed in to change notification settings

monomest/Kaldi_CU

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

35 Commits
 
 
 
 
 
 

Repository files navigation

Kaldi_CU

Kaldi recipe for CU Kids Speech Corpus data model building

Steps: Make sure kaldi is installed first before proceeding.

  1. Download the CU Kids Speech Corpus dataset
  2. In s5/run.sh modify CU_ROOT to point to the main directory of the dataset
  3. ./run.sh This will build the HMM-GMM model
  4. After completed successfully run local/nnet3/run_tdnn_delta.sh for the TDNN model THe script s5/clean.sh will remove file created from s5/run.sh so that you can train the models again.

Install Kaldi: Refer to: https://kaldi-asr.org/doc/tutorial_setup.html

  1. git clone https://github.com/kaldi-asr/kaldi.git
  2. Look at the kaldi/INSTALL file and follow the instructions there
  3. Download SRILM by running kaldi/tools/install_srilm.sh

Initialising katana:

To execute the steps using the supercomputer katana.

  1. ssh [email protected] in terminal. Or, use the alias katana.
  2. Create a new screen using screen -S nameOfSession.
  3. Request an interactive GPU node using qsub -I -l select=1:ngpus=2:ncpus=16:mem=80gb,walltime=10:00:00. Once the node is ready, you are now in the node. The terminal will show (zID@kxxx), where kxxx is your node.
  4. Now you are inside the screen, and inside the GPU node. Run whatever process you need.
  5. Load modules by running
module load cuda/10.1
module load python/3.6.5
module load intel/19.0.0.117
module load sox/14.4.2

Note: If there is an error message saying Permission Denied when running a script, use chmod u+x -R /path/to/directory to change the permissions of all the files in the directory so that you have permission to execute.

To install any python packages not in katana, use a virtual environment: https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/ The virtual environment I am using currently is kaldi/egs/Kaldi_CU source Kaldi_CU/bin/activate Includes num2words package.

Leaving katana Asumming you are inside a screen, and inside a requested GPU node.

  1. CtrlA D to detach from the screen session.
  2. exit to logout of the katana session.

Returning to katana

  1. ssh [email protected] in terminal. Or, use the alias katana.
  2. Go back to your screen screen -r nameOfSession eg. screen -r ogi

Useful katana screen things

  • To create a new window (tab) within a screen, use CtrlA C
  • To go to next and previous windows, use CtrlA N and CtrlA P respectively.
  • To check if you are in a screen, type the command echo $TERM

About

Scripts for training acoustic and language models using CU Kids' Speech Corpus. Also includes the data preparation scripts.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published