Deep Neural Network Embedding For Text-Independent Speaker Verification

This text is study for danielpovey.com/files/2017_interspeech_embeddings.pdf

Overview

• Replace i-vectors to embedding extracted from feed-forward deep neural network for text-independent speaker verification
• Use temporal pooling layer to aggregate variable length of speech segments

DNN embedding system

• Features
  • 20-dimensional MFCC with 25ms frame-length
  • Energy-based VAD is applied
• Model: Time-delay neural network (TDNN)
  • The first 5 layers work at the frame level
  • The statistic pooling layer aggregates over the input segment and compute its mean and standard deviation
  • Two fully connected layer control dimension to 512 and 300 (number of speaker)
  • Total parameter: 4.4M
  • Use embedding a and b
• PLDA backend
  • Embedding is reduced using LDA
  • Embedding length normalization is applied 
  • PLDA scores are normalized using adative s-norm

Experiments

• Training data
  • SWBD
    • 65,000 recordings
    • 6,500 speakers
• Evaluation
  • Asses performance using NIST 2010 (SRE10), 2016 (SRE16)
    • SRE10
      • Consist of English telephone speech
      • Enrollment utterances are full-length
      • Test utterances have been truncated to the first T seconds of speech
    • SRE16
      • Consist of Tagalog and Cantonese language telephone speech
      • Entollment utterances contain about 60seconds of speech
      • Test utterances range from 10 to 60 seconds
  • Results

• For short length of test set, proposed model shows better results
• Fusion system shows the best results

Conclusions

• Propose DNN-based frame-level feature extraction for text-dependent speaker verification
• Overall, the embeddings seem to be competitive with traditional i-vector
• For short utterance, DNN-based feature shows better performance