Speaker-independent Recognizer Research Articles

Atributos eficientes em reconhecimento automático de voz distribuído

Este artigo descreve e examina atributos eficientes para sistemas de reconhecimento automático de voz (RAV) distribuídos. Os atributos são extraídos a partir de transformações dos parâmetros dos codificadores utilizados em redes móveis celulares e em redes IP que são as principais aplicações visadas em reconhecimento de voz distribuído. Em particular, são analisados atributos obtidos de parâmetros LPC e LSF, em intervalos de 10 ms e 20 ms, utilizando um reconhecedor independente do locutor baseado em HMM. É examinado também o melhor domínio para se realizar a interpolação dos parâmetros em reconhecimento automático de voz distribuído.

Speaker verification method and apparatus using mixture decomposition discrimination

A new speaker verification method, termed Mixture Decomposition Discrimination, (MDD) and a new apparatus for using MDD are presented. MDD takes mixture component score information from a speaker independent recognizer and transmits this information while it is still decomposed as a mixture of component scores that indicate the response of the states of the HMM before this information is combined into a single speaker independent recognizer parameter. MDD can be very effective in improving the performance of existing verification methods based on speaker dependent HMMs with cohort normalization because the errors of the two speaker verification methods are very uncorrelated statistically. Experimental results have shown that when MDD is incorporated into a system that also uses speaker dependent HMMs, the resulting hybrid system has its average equal error rate reduced by 46% compared to cohort normalized speaker independent HMM. MDD is used with a speaker dependent linear discriminator which has relatively low computational and storage requirements. Thus, the increased performance of a hybrid MDD/CNHMM system may be achieved with minimal increase in computational and data storage assets.

Energy conditioned spectral estimation for recognition of noisy speech

An estimation algorithm to improve the noise robustness of filterbank-based speech recognition systems is presented. The algorithm is based on a minimum mean square error (MMSE) estimation of the filter log-energies, introducing a significant improvement over related published algorithms by conditioning the estimate on the total frame energy. The algorithm was evaluated with DECIPHER, SRI's continuous-speech speaker-independent recognizer, on two types of noisy speech: a standard database with added white Gaussian noise, and recordings made in a noisy environment. With white noise the recognition accuracy obtained while training on clean speech and testing in noise approached that obtained with training and testing in noise. In the noisy environment, the estimation algorithm boosted the recognition system's performance with a table mounted microphone almost to the level achieved with a close talking microphone.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

A new approach to connected digit recognition

A new approach to connected digit recognition is described. The recognition system is based on a novel dynamic time warping (DTW) algorithm which efficiently builds candidate strings from successive reference patterns of isolated words. The system operates over dialed‐up telephone lines and can be used as either a speaker trained or as a speaker independent recognizer. The DTW algorithm is described and the effects on performance of several key parameters of the recognition system are discussed. An important aspect of this connected digit recognizer is the effect of the training procedure on the performance of the system. It will be shown that when isolated word reference templates are used, it is important to perform appropriate normalization of the length of the reference templates to achieve good performance scores. In formal evaluation tests, average digit accuracy of about 99% with average string accuracy of about 96% was obtained with 480 variable length strings (from six talkers) for both speaker trained and speaker independent templates.