Command Word Recognition is based on a phoneme level GMM-HMM embedding model

pip install -r requirements.txt

Build lexicon which maps a character to phonemes

python DaCiDian/DaCiDian.py

python prep_data.py --mfcc_dim 13 --min_thresh 5

phoneme whose frequency less than 5 will be replaced by 'UNK'

python main.py

it takes about 4 min for a step on my Win10 and about 45s on my Manjaro Linux. Default GMM feature dim is 13. It can give better acc on test set than dim=39

python test.py

Remember to change model_path in test.py. It will give viterbi align and accuracy.

results of 13 dim mfcc feature:

results of 39 dim mfcc feature:

viterbi align example:

• DaCiDian
• hmmlearn
• https://zhuanlan.zhihu.com/p/55826713
• https://github.com/desh2608/gmm-hmm-asr/blob/master/submission.py
• https://blog.csdn.net/chinatelecom08/article/details/82901480