为什么深度神经网络适合语音识别.pdf

Why Deep Neural Network Works for Speech Recognition? Hui Jiang Department of Electrical Engineering and Computer Science Lassonde School of Engineering, York University, CANADA Joint work with Y. Bao, J. Pan, O. Abdel-Hamid Outline •  Introduction o  Automatic Speech Recognition (ASR) o  Deep Neural Network (DNN) •  DNN/HMM for Speech •  Bottleneck Features •  Incoherent Training •  Conclusions Introduction: ASR History •  ASR formulation: o  GMM/HMM + n-gram + Viterbi search •  Technical advances (incremental) over past 10 years: o  Adaptation (speaker/environment): 5% rel. gain o  Discriminative Training: 5-10% rel. gain o  Feature normalization: 5% rel. gain o  ROVER: 5% rel. gain •  More and more data  better and better accuracy o  read speech (90%), telephony speech (70%) o  meeting/voicemail recording (60%) Acoustic Modeling: Optimization •  Acoustic modeling  large-scale optimization o  2000+ hour data  GMMs/HMM billions of samples  10+ million free parameters •  Training Methods o  Maximum Likelihood Estimation (MLE) o  Discriminative Training (DT) •  Engineering Issues o  Efficiency: feasible with 100-1000 of CPUs o  Reliability: robust estimation of all parameters Neural Network for ASR •  1990s: MLP for ASR (Bourlard and Morgan, 1994) o  NN/HMM hybrid model (worse than GMM/HMM) •  2000s: TANDEM (Hermansky, Ellis, et al., 2000) o  Use MLP as Feature Extraction (5-10% rel. gain) •  2006: DNN for small tasks (Hinton et al., 2006) o  RBM-based pre-training for DNN •  2010: DNN for small-scale ASR (Mohamed, Yu, et al. 2010) •  2011: DNN for large-scale ASR o  Over 30% rel. gain in Switchboard (Seide et al., 2011) Deep Neural Network (DNN) DNN Training (I) Given a