Connectionist probability estimators in HMM speech recognition

Abstract

We are concerned with integrating connectionist networks into a hidden Markov model (HMM) speech recognition system. This is achieved through a statistical interpretation of connectionist networks as probability estimators. We review the basis of HMM speech recognition and point out the possible benefits of incorporating connectionist networks. Issues necessary to the construction of a connectionist HMM recognition system are discussed, including choice of connectionist probability estimator. We describe the performance of such a system using a multilayer perceptron probability estimator evaluated on the speaker-independent DARPA Resource Management database. In conclusion, we show that a connectionist component improves a state-of-the-art HMM system.

Investigation of maxout networks for speech recognition

“…Neural network acoustic models (AMs) [1,2] and its recent resurgence in the form of deep neural networks (DNNs) [3] have been successfully applied across a range of different automatic speech recognition (ASR) tasks. Some contemporary examples include i) initial work on phone classification [4], ii) conversational-style large vocabulary speech recognition (LVSR) systems [5,6,7,8], iii) noise robust applications [9], iv) various aspects of multi-and cross-lingual learning schemes [10,11,12,13,14] and v) distant and multichannel LVSR of meetings [15].…”

Impact of single-microphone dereverberation on DNN-based meeting transcription systems

“…Of these techniques, this work concerns front-end approaches, which attempt to eliminate the effect of distortion from feature vectors. It is sometimes argued that front-end approaches are advantageous since they can be used with any forms of acoustic models, including those based on deep neural networks (DNNs) [1,2], which have recently been breaking records in several speech recognition tasks. Nevertheless, these techniques have often been evaluated by using classical but less sophisticated systems and small tasks with a significant degree of mismatch between training and testing environments.…”

“…Their work achieved the best published result for the Aurora4 multi-condition task, which is an artificially created medium vocabulary task. They also showed that preprocessing feature vectors with a cepstral-domain MMSE noise reduction algorithm [8] degraded the performance of the DNN-based acoustic model 1 . These results clearly indicate that a significant portion of the gains from existing environmental robustness techniques might not be carried over to state-of-the-art systems in practical scenarios, calling for further investigation.…”

Speaker Dependent and Independent Isolated Hindi Word Recognizer using Hidden Markov Model (HMM)

“…The reason behind choosing the Hidden Markov Model over other existing techniques; for example DTW, ANN etc., is that HMM has been proven to be one of the most powerful and dominating statistical approach over the last past years and have been applied in a no. of models like pattern [1,2], DNA sequence analysis [3,4], pathologies [5] or speech recognition [6,7]. The basic theory of HMM was published in a series of classic papers by Baum and his colleagues in the late 1960s and early 1970s which was then implemented for speech recognition applications by Baker at Carnegie Mellon University (CMU) and by Jelinek and his colleagues at IBM in the 1970s [8][9][10][11][12][13][14][15][16][17][18][19][20] and further explored by L. Rabiner, et al [21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36] in 1980s and the early 1990s.…”