Masahide Sugiyama

Noise-robust HMMs based on minimum error classification

The authors compare and contrast the noise-robustness of hidden Markov models (HMMs) trained using a discriminant minimum error classification (MEC) optimization criterion with that of HMMs trained using the conventional maximum likelihood (ML) approach. Isolated word recognition experiments were performed on the ATR 5240 Japanese word database. MEC continuous Gaussian mixture density HMMs trained in a specific noisy environment were found to be more robust to changes in the signal-to-noise ratio than conventional ML HMMs. MEC HMMs trained in various noisy environments were more robust in all environments than conventional ML HMMs.<<ETX>>

A segment-based speaker adaptation neural network applied to continuous speech recognition

The authors describe a speaker adaptation technique using segment-based neural-mapping applied to continuous speech recognition. The adaptation neural network has a time-shifted subconnection architecture to maintain the temporal structure in the acoustic segment and to decrease the amount of speech data for training. The effectiveness of this network has been reported for phoneme recognition. The speaker adaptation network is combined with a TDNN-LR continuous speech recognizer, and is evaluated in word and phrase recognition experiments with several speakers. The results of 500-word recognition experiments show that the recognition rate by segment-based adaptation is 92.2%, 28.8% higher than the rate without adaptation. The results of 278 phrase recognition experiments show that the recognition rate by segment-based adaptation is 57.4%, 27.7% higher than the rate without adaptation.<<ETX>>

Speaker-independent phoneme recognition using large-scale neural networks

The authors describe a large-scale neural network architecture based on TDNN (time-delay neural networks) for speaker-independent phoneme recognition which represents an advance over speaker-dependent and multi-speaker phoneme recognition. Based on a preliminary study on speaker-independent phoneme recognition for voiced stops mod b,d,g mod , a large-scale network is constructed with about 330000 connections in a modular fashion. For speaker-independent all-consonant recognition, a multi-speaker training approach is implemented with several devices in the process of training. This network finally achieved favorable results for speaker-independent phoneme recognition.<<ETX>>