ASR_2018_T01

Data pre processing

Run python import_timit.py --timit=./TIMIT --preprocessed=False to compute the features and store them in a folder. This script also converts the NIST "SPHERE" file format to WAVE PCM format. If you have already converted the files, set --preprocessed=True to skip the conversion process.

References:

• Mel Frequency Cepstral Coefficient (MFCC) tutorial :
  • http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/
• TIMIT related documents:
  • https://nvlpubs.nist.gov/nistpubs/Legacy/IR/nistir4930.pdf
  • https://github.com/philipperemy/timit
• Implementation references:
  • http://scikit-learn.org/stable/index.html
  • http://scikit-learn.org/stable/modules/generated/sklearn.mixture.GaussianMixture.html
  • http://www.pitt.edu/~naraehan/python2/pickling.html
  • https://github.com/belambert/asr-evaluation

This repo contains

• Code to read files and compute MFCC features
• Computing MFCC for time slices given in .PHN files
• Dumping computed features to a folder
• Dumping phone-wise features to a folder
• GMM training
• GMM model dumping
• GMM evaluation
• PER computation

Expected:

- Implement and run train.py and test.py
- Please feel free to modify any part of the code to implement the solution