Regularized subspace Gaussian mixture models for cross-lingual 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.

Unsupervised cross-lingual knowledge transfer in DNN-based LVCSR

“…Before discussing the results on GlobalPhone, it is important to note that the results reported in various sources (for example, [1,3,7,23]) are not directly comparable. This primarily because of the differences between LMs, which are much more significant than other differences, such as the use of MFCC vs PLP features.…”

“…Our setup is similar to that reported in [7]. We use German as our in-domain language and we simulate different degrees of available resources by selecting random 1 and 5 hour subsets of the total 15 hours of labeled training speech data.…”

“…A number of techniques for cross-lingual acoustic modelling have been published including the use of: global phone sets [1,2]; multilingual posterior features in tandem [3,4] and Kullback-Liebler hidden Markov model (KL-HMM) systems [5]; subspace Gaussian mixture models (SGMMs) with a shared multilingual phonetic subspace [6,7]; and crosslingual bootstrapping with unsupervised training of the target language [8]. These approaches rely on transcribed audio data for building automatic speech recognition (ASR) systems in some source languages that may or may not be linguistically related to the target language.…”

Deep neural network features and semi-supervised training for low resource speech recognition

“…In these settings, it is possible to take advantage of transcribed data from other languages to build multilingual acoustic models [1,2]. Multilingual training with Subspace Gaussian Mixture Models [3] have also been proposed to train acoustic models [4,5].…”

Joint Uncertainty Decoding for Noise Robust Subspace Gaussian Mixture Models

“…This allows a much larger number of Gaussian components to be used by each HMM state while the total number of parameters to be estimated is typically smaller compared to conventional GMM-based acoustic models. Recent research has indicated that an SGMM acoustic model may result in more accurate speech recognition compared with its GMM counterpart, in both monolingual and multilingual settings [17]- [21]. However, in noisy environments uncompensated SGMMs suffer similar problems to conventional GMMs.…”