Recently, developing Automatic Speech Recognition (ASR) systems for Low Resource (LR) languages is an active research area. The research in ASR is significantly advanced using deep learning approaches producing state-of-the-art results compared to the conventional approaches. However, it is still challenging to use such approaches for LR languages since it requires a huge amount of training data. Recently, data augmentation, multilingual and cross-lingual approaches, transfer learning, etc. enable training deep learning architectures. This paper presents an overview of deep learning-based approaches for building ASR for LR languages. Recent projects and events organized to support the development of ASR and related applications in this direction are also discussed. This paper could be a good motivation for the researchers interested to work towards low resource ASR using deep learning techniques. The approaches described here could be useful in other related applications, such as audio search.
{"title":"Advances in Low Resource ASR: A Deep Learning Perspective","authors":"Hardik B. Sailor, Ankur T. Patil, H. Patil","doi":"10.21437/SLTU.2018-4","DOIUrl":"https://doi.org/10.21437/SLTU.2018-4","url":null,"abstract":"Recently, developing Automatic Speech Recognition (ASR) systems for Low Resource (LR) languages is an active research area. The research in ASR is significantly advanced using deep learning approaches producing state-of-the-art results compared to the conventional approaches. However, it is still challenging to use such approaches for LR languages since it requires a huge amount of training data. Recently, data augmentation, multilingual and cross-lingual approaches, transfer learning, etc. enable training deep learning architectures. This paper presents an overview of deep learning-based approaches for building ASR for LR languages. Recent projects and events organized to support the development of ASR and related applications in this direction are also discussed. This paper could be a good motivation for the researchers interested to work towards low resource ASR using deep learning techniques. The approaches described here could be useful in other related applications, such as audio search.","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"33 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129740864","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Mining Training Data for Language Modeling Across the World's Languages","authors":"Manasa Prasad, Theresa Breiner, D. Esch","doi":"10.21437/SLTU.2018-13","DOIUrl":"https://doi.org/10.21437/SLTU.2018-13","url":null,"abstract":"","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"28 2 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133487657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Inspired by infant language acquisition, unsupervised subword discovery of zero-resource languages has gained attention recently. The Dirichlet Process Gaussian Mixture Model (DPGMM) achieves top results evaluated by the ABX discrimination test. However, the DPGMM model is too sensitive to acoustic variation and often produces too many types of subword units and a relatively high-dimensional posteriorgram, which implies high computational cost to perform learning and inference, as well as more tendency to be overfitting. This paper proposes applying functional load to reduce the number of sub-word units from DPGMM. We greedily merge pairs of units with the lowest functional load, causing the least information loss of the language. Results on the Xitsonga corpus with the official setting of Zerospeech 2015 show that we can reduce the number of sub-word units by more than two thirds without hurting the ABX error rate. The number of units is close to that of phonemes in human language.
{"title":"Optimizing DPGMM Clustering in Zero Resource Setting Based on Functional Load","authors":"Bin Wu, S. Sakti, Jinsong Zhang, Satoshi Nakamura","doi":"10.21437/SLTU.2018-1","DOIUrl":"https://doi.org/10.21437/SLTU.2018-1","url":null,"abstract":"Inspired by infant language acquisition, unsupervised subword discovery of zero-resource languages has gained attention recently. The Dirichlet Process Gaussian Mixture Model (DPGMM) achieves top results evaluated by the ABX discrimination test. However, the DPGMM model is too sensitive to acoustic variation and often produces too many types of subword units and a relatively high-dimensional posteriorgram, which implies high computational cost to perform learning and inference, as well as more tendency to be overfitting. This paper proposes applying functional load to reduce the number of sub-word units from DPGMM. We greedily merge pairs of units with the lowest functional load, causing the least information loss of the language. Results on the Xitsonga corpus with the official setting of Zerospeech 2015 show that we can reduce the number of sub-word units by more than two thirds without hurting the ABX error rate. The number of units is close to that of phonemes in human language.","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115221893","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Development of Assamese Continuous Speech Recognition System","authors":"Tanmay Bhowmik, S. Mandal","doi":"10.21437/SLTU.2018-46","DOIUrl":"https://doi.org/10.21437/SLTU.2018-46","url":null,"abstract":"","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"569 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123322906","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Analysis and Comparison of Features for Text-Independent Bengali Speaker Recognition","authors":"S. Das, P. Das","doi":"10.21437/SLTU.2018-57","DOIUrl":"https://doi.org/10.21437/SLTU.2018-57","url":null,"abstract":"","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128985535","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Improved Language Identification Using Stacked SDC Features and Residual Neural Network","authors":"R. Vuddagiri, Hari Krishna Vydana, A. Vuppala","doi":"10.21437/SLTU.2018-44","DOIUrl":"https://doi.org/10.21437/SLTU.2018-44","url":null,"abstract":"","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125326657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Building accurate acoustic models for low resource languages is the focus of this paper. Acoustic models are likely to be accurate provided the phone boundaries are determined accurately. Conventional flat-start based Viterbi phone alignment (where only utterance level transcriptions are available) results in poor phone boundaries as the boundaries are not explicitly modeled in any statistical machine learning system. The focus of the effort in this paper is to explicitly model phrase boundaries using acoustic cues obtained using signal processing. A phrase is made up of a sequence of words, where each word is made up of a sequence of syllables. Syllable boundaries are detected using signal processing. The waveform corresponding to an utterance is spliced at phrase boundaries when it matches a syllable boundary. Gaussian mixture model - hidden Markov model (GMM-HMM) training is performed phrase by phrase, rather than utterance by utterance. Training using these short phrases yields better acoustic models. This alignment is then fed to a DNN to enable better discrimination between phones. During the training process, the syllable boundaries (obtained using signal processing) are restored in every iteration. A rela-tive improvement is observed in WER over the baseline Indian languages, namely, Gujarati, Tamil, and Telugu.
{"title":"Signal Processing Cues to Improve Automatic Speech Recognition for Low Resource Indian Languages","authors":"Arun Baby, S. KarthikPandiaD., H. Murthy","doi":"10.21437/SLTU.2018-6","DOIUrl":"https://doi.org/10.21437/SLTU.2018-6","url":null,"abstract":"Building accurate acoustic models for low resource languages is the focus of this paper. Acoustic models are likely to be accurate provided the phone boundaries are determined accurately. Conventional flat-start based Viterbi phone alignment (where only utterance level transcriptions are available) results in poor phone boundaries as the boundaries are not explicitly modeled in any statistical machine learning system. The focus of the effort in this paper is to explicitly model phrase boundaries using acoustic cues obtained using signal processing. A phrase is made up of a sequence of words, where each word is made up of a sequence of syllables. Syllable boundaries are detected using signal processing. The waveform corresponding to an utterance is spliced at phrase boundaries when it matches a syllable boundary. Gaussian mixture model - hidden Markov model (GMM-HMM) training is performed phrase by phrase, rather than utterance by utterance. Training using these short phrases yields better acoustic models. This alignment is then fed to a DNN to enable better discrimination between phones. During the training process, the syllable boundaries (obtained using signal processing) are restored in every iteration. A rela-tive improvement is observed in WER over the baseline Indian languages, namely, Gujarati, Tamil, and Telugu.","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128106054","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"JAMLIT: A Corpus of Jamaican Standard English for Automatic Speech Recognition of Children's Speech","authors":"Stefan Watson, André Coy","doi":"10.21437/SLTU.2018-51","DOIUrl":"https://doi.org/10.21437/SLTU.2018-51","url":null,"abstract":"","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133130077","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The indulgent acquaintance of mathematical basic concepts creates the pavement for numerous opportunities in life for every individual, including visually impaired people. The use of assertive technology for the disabled section of the society makes them more independent and avoid barriers in the field of education and employment. This research is focused to design an Android-based application i.e. talking Calculator for low resource based Marathi native language. The novelty of this work is to develop both, the application and the Marathi number corpus. Marathi is an Indo-Aryan language spoken by approximately 6.99 million speakers in India, which is the third widely spoken language after Bengali and Telugu but as they lack in linguistic resources, e.g. grammars, POS taggers, corpora, it falls into the category of low resource languages. The front end part of the application depicts the screen of a basic calculator with numerals displayed in Marathi. During runtime, each number is spoken as the specific key is pressed. It also speaks out the operation which is intended to be performed. The concatenation synthesis technique is applied to speak out the value of decimal places in the output number. The result is spoken out with proper place value of a digit in Marathi. The performance of the system is measured to the accuracy rate of 95.5%. The average run time complexity of the application is also calculated which is noted down to 2.64 sec. The feedback and review of the application is also taken from real end-user i.e. blind people.
{"title":"Implementation of Concatenation Technique for Low Resource Text-To-Speech System Based on Marathi Talking Calculator","authors":"Monica R. Mundada, Sangramsing Kayte, P. Das","doi":"10.21437/SLTU.2018-16","DOIUrl":"https://doi.org/10.21437/SLTU.2018-16","url":null,"abstract":"The indulgent acquaintance of mathematical basic concepts creates the pavement for numerous opportunities in life for every individual, including visually impaired people. The use of assertive technology for the disabled section of the society makes them more independent and avoid barriers in the field of education and employment. This research is focused to design an Android-based application i.e. talking Calculator for low resource based Marathi native language. The novelty of this work is to develop both, the application and the Marathi number corpus. Marathi is an Indo-Aryan language spoken by approximately 6.99 million speakers in India, which is the third widely spoken language after Bengali and Telugu but as they lack in linguistic resources, e.g. grammars, POS taggers, corpora, it falls into the category of low resource languages. The front end part of the application depicts the screen of a basic calculator with numerals displayed in Marathi. During runtime, each number is spoken as the specific key is pressed. It also speaks out the operation which is intended to be performed. The concatenation synthesis technique is applied to speak out the value of decimal places in the output number. The result is spoken out with proper place value of a digit in Marathi. The performance of the system is measured to the accuracy rate of 95.5%. The average run time complexity of the application is also calculated which is noted down to 2.64 sec. The feedback and review of the application is also taken from real end-user i.e. blind people.","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121508920","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this paper, we present a development of Automatic Speech Recognition (ASR) system as a part of a speech-based access for an agricultural commodity in the Gujarati (a low resource) language. We proposed to use neural networks for language modeling, acoustic modeling, and feature learning from the raw speech signals. The speech database of agricultural commodities was collected from the farmers belonging to various villages of Gujarat state (India). The database has various dialectal variations and real noisy acoustic environments. Acoustic modeling is performed using Time Delay Neural Networks (TDNN). The auditory feature representation is learned using Convolutional Restricted Boltzmann Machine (ConvRBM) and Teager Energy Operator (TEO). The language model (LM) rescoring is performed using Recurrent Neural Networks (RNN). RNNLM rescoring provides an absolute reduction of 0.69-1.18 in % WER for all the feature sets compared to the bi-gram LM. The system combination of ConvRBM and Mel filterbank further improved the performance of ASR compared to the baseline TDNN with Mel filterbank features (5.4 % relative reduction in WER). The statistical significance of proposed approach is justified using a bootstrap-based % Probability of Improvement (POI) measure.
{"title":"Neural Networks-based Automatic Speech Recognition for Agricultural Commodity in Gujarati Language","authors":"Hardik B. Sailor, H. Patil","doi":"10.21437/SLTU.2018-34","DOIUrl":"https://doi.org/10.21437/SLTU.2018-34","url":null,"abstract":"In this paper, we present a development of Automatic Speech Recognition (ASR) system as a part of a speech-based access for an agricultural commodity in the Gujarati (a low resource) language. We proposed to use neural networks for language modeling, acoustic modeling, and feature learning from the raw speech signals. The speech database of agricultural commodities was collected from the farmers belonging to various villages of Gujarat state (India). The database has various dialectal variations and real noisy acoustic environments. Acoustic modeling is performed using Time Delay Neural Networks (TDNN). The auditory feature representation is learned using Convolutional Restricted Boltzmann Machine (ConvRBM) and Teager Energy Operator (TEO). The language model (LM) rescoring is performed using Recurrent Neural Networks (RNN). RNNLM rescoring provides an absolute reduction of 0.69-1.18 in % WER for all the feature sets compared to the bi-gram LM. The system combination of ConvRBM and Mel filterbank further improved the performance of ASR compared to the baseline TDNN with Mel filterbank features (5.4 % relative reduction in WER). The statistical significance of proposed approach is justified using a bootstrap-based % Probability of Improvement (POI) measure.","PeriodicalId":190269,"journal":{"name":"Workshop on Spoken Language Technologies for Under-resourced Languages","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114635431","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：