Automatic Speech Recognition with HMM

• Feature Extraction:
  • 39 MFCC features
• Acoustic Model:
  • Mixture Gaussian model for
• Lexicon/Pronunciation Model
  • HMM: what phones can follow each other
• Language Model
  • N-grams for computing
• Decoder
  • Viterbi Algorithm: dynamic programming for combining all these to get word squence from speech

To execute:
python main.py
To print results:
python summary.py

Viterbi algorithm ( Max-Product algorithm )

Objective: Find the most probable sequences of states that maximizes the posterior

Study link

• Stanford CS224s

Result example, 10 mfccs: