Regularized Subspace Gaussian Mixture Models for Speech Recognition

Abstract

Abstract-We investigate cross-lingual acoustic modelling for low resource languages using the subspace Gaussian mixture model (SGMM). We assume the presence of acoustic models trained on multiple source languages, and use the global subspace parameters from those models for improved modelling in a target language with limited amounts of transcribed speech. Experiments on the GlobalPhone corpus using Spanish, Portuguese, and Swedish as source languages and German as target language (with 1 hour and 5 hours of transcribed audio) show that multilingually trained SGMM shared parameters result in lower word error rates (WERs) than using those from a single source language. We also show that regularizing the estimation of the SGMM state vectors by penalizing their 1-norm help to overcome numerical instabilities and lead to lower WER.

Cross-Lingual Subspace Gaussian Mixture Models for Low-Resource Speech Recognition

“…After increasing the amount of training data to be 15 hours, we did not obtain improvement by applying the 1 -norm regularization as shown in Figure 6. This agrees with our previous experience of using 1 -norm regularization for SGMMs [30] on a different task.…”

“…This problem is most acute for the state-dependent vectors v jk -unlike the globally shared parameters Φ i , they are only trained on those speech frames which align with the corresponding sub-state. To overcome this problem, we proposed a regularized ML estimate for the state vectors [30] in which penalties based on the 1 -norm and 2 -norm of the state vectors, as well as their linear combination (the elastic net [31]), were investigated. Regularization using the 1 -norm penalty was found to be best suited in cross-lingual settings where the amount of target training data is very limited [28].…”

“…However, the solution of the auxiliary function is not readily available for the 1 -norm penalty, since the derivative of the auxiliary function is not continuous. We have previously applied the gradient projection based optimization approach [35] to obtain the solution [30]. The idea of regularization can also be applied to other types of parameters in SGMMs.…”

Maximum a posteriori adaptation of subspace Gaussian mixture models for cross-lingual speech recognition

“…Hence, the performance differences are purely owing to better parameter estimation. For SGMM systems with S = 40, regularized state vector estimation by 1-norm penalty [10] is applied to improve numerical stability; we have also observed that such regularisation brings gains in accuracy [3]. For comparison, the monolingual GMM and SGMM systems with the entire 14.8 hours of target language training data available in GlobalPhone achieve 25.7% and 24.0% WER 2 .…”

Realce E Reconhecimento De Voz Contínua Em Ambientes Adversos

“…Em [170] foi proposta uma outra alternativa para compensar modelos baseada na mistura de Gaussianas. Essa técnica, chamada subspace gaussian mixture models (SGMM), utiliza um modelo de subespaços globalmente compartilhado entre os estados, de forma a capturar as maiores virações do modelo, provendo uma representação compacta dos modelos acústicos resultando em uma estimativa robusta de atributos, o que permite obter melhoras no desempenho dos sistemas de reconhecimento.…”