Frame erasure concealment is studied to solve the problem of rapid speech quality reduction due to the loss of speech parameters during speech transmission. A large hidden Markov model is applied to model the immittan...Frame erasure concealment is studied to solve the problem of rapid speech quality reduction due to the loss of speech parameters during speech transmission. A large hidden Markov model is applied to model the immittance spectral frequency (ISF) parameters in AMR-WB codec to optimally estimate the lost ISFs based on the minimum mean square error (MMSE) rule. The estimated ISFs are weighted with the ones of their previous neighbors to smooth the speech, resulting in the actual concealed ISF vectors. They are used instead of the lost ISFs in the speech synthesis on the receiver. Comparison is made between the speech concealed by this algorithm and by Annex I of G. 722. 2 specification, and simulation shows that the proposed concealment algorithm can lead to better performance in terms of frequency-weighted spectral distortion and signal-to-noise ratio compared to the baseline method, with an increase of 2.41 dB in signal-to-noise ratio (SNR) and a reduction of 0. 885 dB in frequency-weighted spectral distortion.展开更多
This paper applied Maximum Entropy (ME) model to Pinyin-To-Character (PTC) conversion in-stead of Hidden Markov Model (HMM) that could not include complicated and long-distance lexical informa-tion. Two ME models were...This paper applied Maximum Entropy (ME) model to Pinyin-To-Character (PTC) conversion in-stead of Hidden Markov Model (HMM) that could not include complicated and long-distance lexical informa-tion. Two ME models were built based on simple and complex templates respectively, and the complex one gave better conversion result. Furthermore, conversion trigger pair of y A → y B cBwas proposed to extract the long-distance constrain feature from the corpus; and then Average Mutual Information (AMI) was used to se-lect conversion trigger pair features which were added to the ME model. The experiment shows that conver-sion error of the ME with conversion trigger pairs is reduced by 4% on a small training corpus, comparing with HMM smoothed by absolute smoothing.展开更多
基金The Science Foundation of Southeast University(No.XJ0704268)the Natural Science Foundation of the Education Department of Anhui Province(No.KJ2007B088)
文摘Frame erasure concealment is studied to solve the problem of rapid speech quality reduction due to the loss of speech parameters during speech transmission. A large hidden Markov model is applied to model the immittance spectral frequency (ISF) parameters in AMR-WB codec to optimally estimate the lost ISFs based on the minimum mean square error (MMSE) rule. The estimated ISFs are weighted with the ones of their previous neighbors to smooth the speech, resulting in the actual concealed ISF vectors. They are used instead of the lost ISFs in the speech synthesis on the receiver. Comparison is made between the speech concealed by this algorithm and by Annex I of G. 722. 2 specification, and simulation shows that the proposed concealment algorithm can lead to better performance in terms of frequency-weighted spectral distortion and signal-to-noise ratio compared to the baseline method, with an increase of 2.41 dB in signal-to-noise ratio (SNR) and a reduction of 0. 885 dB in frequency-weighted spectral distortion.
基金Supported by the National Natural Science Foundation of China as key program (No.60435020) and The HighTechnology Research and Development Programme of China (2002AA117010-09).
文摘This paper applied Maximum Entropy (ME) model to Pinyin-To-Character (PTC) conversion in-stead of Hidden Markov Model (HMM) that could not include complicated and long-distance lexical informa-tion. Two ME models were built based on simple and complex templates respectively, and the complex one gave better conversion result. Furthermore, conversion trigger pair of y A → y B cBwas proposed to extract the long-distance constrain feature from the corpus; and then Average Mutual Information (AMI) was used to se-lect conversion trigger pair features which were added to the ME model. The experiment shows that conver-sion error of the ME with conversion trigger pairs is reduced by 4% on a small training corpus, comparing with HMM smoothed by absolute smoothing.