“ASR Status Report 2016-11-28”版本间的差异

• some work about kazak speech recognition
• cnn visualization
• paper reading

• cnn visualization

• Explore the language recognition models including:
• Evaluate the model in the aspect of sentence and frame, find the accuracy is very high.
• Minimize the language model, train it single and joint with speech model, evaluate its result.

• Continue doing the basic explore of joint training.
• Read paper about multi-language recognition models and others.

• fighting with kazak speech recognition system:because the huge size of HCLG.fst the decoding job always make the sever done.

There are several method I have tried

• change the size or word list and corpus this method not worked very well
• prune the LM .And the parameter been used to prune the LM is 2e-7 the size of LM reduce from 290M to 60M but the result about wer is very poor
• I have upload some result about several experiment to CVSS[1]

• there are too much private affairs about myself so the job about visualization last week has been delayed I will try my best to finish it the week

• Learn MFCC extraction mechanism.
• Read kaldi computer-feature code and find how to change MFCC.
• Frequency-weighting based feature extraction.

• Continue replay detection (Freq-Weighting and Freq-Warping).

• Joint-training on SRE and LRE (LRE task). [2]
  • Tdnn is better than LSTM.
  • LRE is a long-term task.
• Briefly overview Interspeech SRE-related papers.
• CSLT-Replay detection.
  • Baseline done (Freq / Mel domain).
  • performance-driven based Freq-Weighting and Freq-Warping --> Yixiang.

• LRE task.
• Replay detection.

• report for Weekly Reading (a brief review of interspeech16), just prepared;
• language scores as decoding mask (1.multiply probability, very bad; 2.add log-softmax, a little bad)
• training with mask failed

• training with shared layers;
• explore single tasks.