[Computer Science] [2018.05] Robust Acoustic Modeling of Reverberation Based on Time Delay Neural Network

This article Carnegie Mellon University (Author: Emanuel Jöbstl) US master's thesis, a total of 73.

This paper studies robust acoustic modeling based on hidden Markov models for speech recognition systems. The focus of this paper is the time delay neural network. We first designed a time-delay neural network model for acoustic modeling, and gave experimental results to prove that our choice of design parameters is correct. Then, we trained the time-delay neural network on the augmented data and compared it with the performance of the traditional fully connected neural network on the reverberation data.

This work investigates robust acoustic modeling for speech recognition systems based on hidden Markov models. The focus of this work is put on time delay neural networks. We first design a time delay neural network model for acoustic modeling and provide empirical results that justify our choice of design parameters. Then, we train the time delay neural network on augmented data, and compare its performance on reverberated data with conventional fully connected neural networks.

1.   引言
   

2. Basic knowledge
3. Related work
4. TDNN acoustic model design
5. Reverberation data evaluation
6. Conclusion
   Appendix Optimal Decoder Parameters

For more exciting articles, please pay attention to the public account: