Articulatory and Stacked Bottleneck Features for Low Resource Speech Recognition

Abstract

In this paper, we discuss the benefits of using articulatory and stacked bottleneck features (SBF) for low resource speech recognition. Articulatory features (AF) which capture the underlying attributes of speech production are found to be robust to channel and speaker variations. However, building an efficient articulatory classifier to extract AF requires an enormous amount of data. In low resource acoustic modeling, we propose to train the bidirectional long short-term memory (BLSTM) articulatory classifier by pooling data from the available low resource Indian languages, namely, Gujarati, Tamil, and Telugu. This is done in the context of Microsoft Indian Language challenge. Similarly, we train a multilingual bottleneck feature extractor and an SBF extractor using the pooled data. To bias, the SBF network towards the target language, a second network in the stacked architecture was trained using the target language alone. The performance of ASR system trained with stand-alone AF is observed to be at par with the multilingual bottleneck features. When the AF and the biased SBF are appended, they are found to outperform the conventional filterbank features in the multilingual deep neural network (DNN) framework and the high-resolution Mel frequency cepstral coefficient (MFCC) features in the time-delayed neural network(TDNN) framework.

Interspeech 2018 Low Resource Automatic Speech Recognition Challenge for Indian Languages

“…The IITM Speech Lab systems [17] used articulatory and stacked bottlneck (SBN) features in the Acoustic Model by creating an bidirectional long short term memory (BLSTM) articulatory feature classifier using the pooled data. The best system with Word Error Rates of 24.29%, 30.33%, 17.9% for Gujarati, Telugu, Tamil respectively was a TDNN-based system with articulatory and SBN features.…”

Frontier Research on Low-Resource Speech Recognition Technology

“…Assamese, Bengali, Kurmanji, Lithuanian, Pashto, Turkish, and Vietnamese in the IARPA Babel [51] BLSTM & TDNN AF and SBF Multilingual corpus assistance and feature stitching Gujarati, Tamil, and Telugu [52] Initially, BN features were regarded as dependent on language, until the speech team of Brno University of Technology proposed a language-independent BN feature extraction framework [39]. The framework is a five-layer MLP with a sigmoid hidden unit, a linear bottleneck, and several output layers.…”

“…Among them, the WER of Gujarati language dropped from 14.61% to 14.11%, and the WER of Telugu language dropped from 21.44% to 19.80%. The team concluded from the experimental results that the proposed combination of AF and SBF is an improvement over traditional features [52].…”

“…Figure 9. (a)The feature extraction model[52]; (b) the pronunciation classifier used to extract AF[52]; and (c) the feature extractor used to extract SBF[52].…”