Acoustic Modeling Based on Deep Learning for Low-Resource Speech Recognition: An Overview

Abstract

The polarization of world languages is becoming more and more obvious. Many languages, mainly endangered languages, are of low-resource attribute due to lack of information. Both language conservation and cultural heritage face important challenges. Therefore, speech recognition for lowresource scenario has become a hot topic in the field of speech. Based on the complex network structures and huge model parameters, deep learning has become a powerful science in the process of speech recognition, which has a broad and far-reaching significance for the study of low-resource speech recognition. Aiming at the characteristic of low resource, this paper reviews the history and research status of two kinds of acoustic models of deep learning neural networks and acoustic end-to-end structures. We further elaborate on several key techniques for improving performance in the two aspects of data and model training. There are two projects for low-resource languages introduced in this paper. The possible future developments are finally pointed out. These works provide some reference for computer speech and language processing.

Vowel classification with combining pitch detection and one‐dimensional convolutional neural network based classifier for gender identification

“…In Ref. [59], DL-based methods, such as a convolutional recurrent neural network (CRNN), temporal convolutional network (TCN), concept-level TCN (CTCN), and CNN, were also applied in both group classification and gender recognition for six categories (as seen in Table 5), which combined different DL models [32,35,37,38] to establish a larger size of the classifier network and offered a gender identification error of <2% and an age group classification error of <20% [59]. In the feature extraction layer, the STFT (with the hamming window) and MFCC methods were used to extract the melscale feature patterns.…”

“…Herein, in classification task, we intend to design a ML‐based or a DL‐based classifier to automatically perform the voice classification and gender identification, including adult males, adult females, and children (boys and girls) [6, 35–37]. To deal with the one‐dimensional (1D) signals, 1D CNN and two‐dimensional (2D) CNN models can be used for digital signal classification in audio and bio‐signal recognition [24, 38–42].…”

“…Herein, in classification task, we intend to design a MLbased or a DL-based classifier to automatically perform the voice classification and gender identification, including adult males, adult females, and children (boys and girls) [6,[35][36][37].…”

Improving Hybrid CTC/Attention Architecture for Agglutinative Language Speech Recognition

“…Orken Zh et al proposed a joint model based on CTC and the attention mechanism for recognition of Kazakh speech in noisy conditions [ 54 ]. In addition to the improvement of the model structure, some important technologies are often applied to low-resource speech recognition, which is also the key to improving performance [ 55 ]. The most widespread application for these is data augmentation, a technology for increasing the amount of data needed for training speech recognition systems.…”

Automatic Speech Recognition Method Based on Deep Learning Approaches for Uzbek Language

“…Several voice assistants can currently recognize human speech patterns through an interactive real-time smart dialogue and apply automatic techniques based on the recognized content, such as Google’s Assistant and Apple’s Siri, which can converse in over 40 and 35 languages, respectively [ 11 ]. The majority of popular ASR systems use Gaussian Mixture Models (GMMs), Hidden Markov Models (HMMs), and Deep Neural Networks (DNNs) [ 12 , 13 , 14 , 15 ]. DNNs play an essential part in the building of ASR systems [ 16 , 17 ], mostly because of the evolution of unique neural network models, as well as training and classification techniques [ 18 , 19 ].…”